diff options
Diffstat (limited to 'webrtc/modules/audio_processing/vad')
29 files changed, 2517 insertions, 0 deletions
diff --git a/webrtc/modules/audio_processing/vad/Android.mk b/webrtc/modules/audio_processing/vad/Android.mk new file mode 100644 index 0000000000..98909b0dda --- /dev/null +++ b/webrtc/modules/audio_processing/vad/Android.mk @@ -0,0 +1,52 @@ +# Copyright (c) 2015 The WebRTC project authors. All Rights Reserved. +# +# Use of this source code is governed by a BSD-style license +# that can be found in the LICENSE file in the root of the source +# tree. An additional intellectual property rights grant can be found +# in the file PATENTS. All contributing project authors may +# be found in the AUTHORS file in the root of the source tree. + +LOCAL_PATH := $(call my-dir) + +include $(CLEAR_VARS) + +include $(LOCAL_PATH)/../../../../android-webrtc.mk + +LOCAL_ARM_MODE := arm +LOCAL_MODULE_CLASS := STATIC_LIBRARIES +LOCAL_MODULE := libwebrtc_apvad +LOCAL_MODULE_TAGS := optional +LOCAL_CPP_EXTENSION := .cc +LOCAL_SRC_FILES := \ + gmm.cc \ + pitch_based_vad.cc \ + pitch_internal.cc \ + pole_zero_filter.cc \ + standalone_vad.cc \ + vad_audio_proc.cc \ + vad_circular_buffer.cc \ + voice_activity_detector.cc \ + +# Flags passed to both C and C++ files. +LOCAL_CFLAGS := \ + $(MY_WEBRTC_COMMON_DEFS) + +LOCAL_CFLAGS_arm := $(MY_WEBRTC_COMMON_DEFS_arm) +LOCAL_CFLAGS_x86 := $(MY_WEBRTC_COMMON_DEFS_x86) +LOCAL_CFLAGS_mips := $(MY_WEBRTC_COMMON_DEFS_mips) +LOCAL_CFLAGS_arm64 := $(MY_WEBRTC_COMMON_DEFS_arm64) +LOCAL_CFLAGS_x86_64 := $(MY_WEBRTC_COMMON_DEFS_x86_64) +LOCAL_CFLAGS_mips64 := $(MY_WEBRTC_COMMON_DEFS_mips64) + +# Include paths placed before CFLAGS/CPPFLAGS +LOCAL_C_INCLUDES := \ + $(LOCAL_PATH) \ + $(LOCAL_PATH)/../../../.. \ + +ifdef WEBRTC_STL +LOCAL_NDK_STL_VARIANT := $(WEBRTC_STL) +LOCAL_SDK_VERSION := 14 +LOCAL_MODULE := $(LOCAL_MODULE)_$(WEBRTC_STL) +endif + +include $(BUILD_STATIC_LIBRARY) diff --git a/webrtc/modules/audio_processing/vad/common.h b/webrtc/modules/audio_processing/vad/common.h new file mode 100644 index 0000000000..be99c1c59d --- /dev/null +++ b/webrtc/modules/audio_processing/vad/common.h @@ -0,0 +1,27 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_VAD_COMMON_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_VAD_COMMON_H_ + +static const int kSampleRateHz = 16000; +static const size_t kLength10Ms = kSampleRateHz / 100; +static const size_t kMaxNumFrames = 4; + +struct AudioFeatures { + double log_pitch_gain[kMaxNumFrames]; + double pitch_lag_hz[kMaxNumFrames]; + double spectral_peak[kMaxNumFrames]; + double rms[kMaxNumFrames]; + size_t num_frames; + bool silence; +}; + +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_VAD_COMMON_H_ diff --git a/webrtc/modules/audio_processing/vad/gmm.cc b/webrtc/modules/audio_processing/vad/gmm.cc new file mode 100644 index 0000000000..9651975913 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/gmm.cc @@ -0,0 +1,64 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/gmm.h" + +#include <math.h> +#include <stdlib.h> + +#include "webrtc/typedefs.h" + +namespace webrtc { + +static const int kMaxDimension = 10; + +static void RemoveMean(const double* in, + const double* mean_vec, + int dimension, + double* out) { + for (int n = 0; n < dimension; ++n) + out[n] = in[n] - mean_vec[n]; +} + +static double ComputeExponent(const double* in, + const double* covar_inv, + int dimension) { + double q = 0; + for (int i = 0; i < dimension; ++i) { + double v = 0; + for (int j = 0; j < dimension; j++) + v += (*covar_inv++) * in[j]; + q += v * in[i]; + } + q *= -0.5; + return q; +} + +double EvaluateGmm(const double* x, const GmmParameters& gmm_parameters) { + if (gmm_parameters.dimension > kMaxDimension) { + return -1; // This is invalid pdf so the caller can check this. + } + double f = 0; + double v[kMaxDimension]; + const double* mean_vec = gmm_parameters.mean; + const double* covar_inv = gmm_parameters.covar_inverse; + + for (int n = 0; n < gmm_parameters.num_mixtures; n++) { + RemoveMean(x, mean_vec, gmm_parameters.dimension, v); + double q = ComputeExponent(v, covar_inv, gmm_parameters.dimension) + + gmm_parameters.weight[n]; + f += exp(q); + mean_vec += gmm_parameters.dimension; + covar_inv += gmm_parameters.dimension * gmm_parameters.dimension; + } + return f; +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/gmm.h b/webrtc/modules/audio_processing/vad/gmm.h new file mode 100644 index 0000000000..9f3e578fef --- /dev/null +++ b/webrtc/modules/audio_processing/vad/gmm.h @@ -0,0 +1,45 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_VAD_GMM_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_VAD_GMM_H_ + +namespace webrtc { + +// A structure that specifies a GMM. +// A GMM is formulated as +// f(x) = w[0] * mixture[0] + w[1] * mixture[1] + ... + +// w[num_mixtures - 1] * mixture[num_mixtures - 1]; +// Where a 'mixture' is a Gaussian density. + +struct GmmParameters { + // weight[n] = log(w[n]) - |dimension|/2 * log(2*pi) - 1/2 * log(det(cov[n])); + // where cov[n] is the covariance matrix of mixture n; + const double* weight; + // pointer to the first element of a |num_mixtures|x|dimension| matrix + // where kth row is the mean of the kth mixture. + const double* mean; + // pointer to the first element of a |num_mixtures|x|dimension|x|dimension| + // 3D-matrix, where the kth 2D-matrix is the inverse of the covariance + // matrix of the kth mixture. + const double* covar_inverse; + // Dimensionality of the mixtures. + int dimension; + // number of the mixtures. + int num_mixtures; +}; + +// Evaluate the given GMM, according to |gmm_parameters|, at the given point +// |x|. If the dimensionality of the given GMM is larger that the maximum +// acceptable dimension by the following function -1 is returned. +double EvaluateGmm(const double* x, const GmmParameters& gmm_parameters); + +} // namespace webrtc +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_VAD_GMM_H_ diff --git a/webrtc/modules/audio_processing/vad/gmm_unittest.cc b/webrtc/modules/audio_processing/vad/gmm_unittest.cc new file mode 100644 index 0000000000..f8e1bde776 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/gmm_unittest.cc @@ -0,0 +1,65 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/gmm.h" + +#include <math.h> + +#include "testing/gtest/include/gtest/gtest.h" +#include "webrtc/modules/audio_processing/vad/noise_gmm_tables.h" +#include "webrtc/modules/audio_processing/vad/voice_gmm_tables.h" + +namespace webrtc { + +TEST(GmmTest, EvaluateGmm) { + GmmParameters noise_gmm; + GmmParameters voice_gmm; + + // Setup noise GMM. + noise_gmm.dimension = kNoiseGmmDim; + noise_gmm.num_mixtures = kNoiseGmmNumMixtures; + noise_gmm.weight = kNoiseGmmWeights; + noise_gmm.mean = &kNoiseGmmMean[0][0]; + noise_gmm.covar_inverse = &kNoiseGmmCovarInverse[0][0][0]; + + // Setup voice GMM. + voice_gmm.dimension = kVoiceGmmDim; + voice_gmm.num_mixtures = kVoiceGmmNumMixtures; + voice_gmm.weight = kVoiceGmmWeights; + voice_gmm.mean = &kVoiceGmmMean[0][0]; + voice_gmm.covar_inverse = &kVoiceGmmCovarInverse[0][0][0]; + + // Test vectors. These are the mean of the GMM means. + const double kXVoice[kVoiceGmmDim] = { + -1.35893162459863, 602.862491970368, 178.022069191324}; + const double kXNoise[kNoiseGmmDim] = { + -2.33443722724409, 2827.97828765184, 141.114178166812}; + + // Expected pdf values. These values are computed in MATLAB using EvalGmm.m + const double kPdfNoise = 1.88904409403101e-07; + const double kPdfVoice = 1.30453996982266e-06; + + // Relative error should be smaller that the following value. + const double kAcceptedRelativeErr = 1e-10; + + // Test Voice. + double pdf = EvaluateGmm(kXVoice, voice_gmm); + EXPECT_GT(pdf, 0); + double relative_error = fabs(pdf - kPdfVoice) / kPdfVoice; + EXPECT_LE(relative_error, kAcceptedRelativeErr); + + // Test Noise. + pdf = EvaluateGmm(kXNoise, noise_gmm); + EXPECT_GT(pdf, 0); + relative_error = fabs(pdf - kPdfNoise) / kPdfNoise; + EXPECT_LE(relative_error, kAcceptedRelativeErr); +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/noise_gmm_tables.h b/webrtc/modules/audio_processing/vad/noise_gmm_tables.h new file mode 100644 index 0000000000..293af57a2a --- /dev/null +++ b/webrtc/modules/audio_processing/vad/noise_gmm_tables.h @@ -0,0 +1,85 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +// GMM tables for inactive segments. Generated by MakeGmmTables.m. + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_VAD_NOISE_GMM_TABLES_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_VAD_NOISE_GMM_TABLES_H_ + +static const int kNoiseGmmNumMixtures = 12; +static const int kNoiseGmmDim = 3; + +static const double + kNoiseGmmCovarInverse[kNoiseGmmNumMixtures][kNoiseGmmDim][kNoiseGmmDim] = { + {{7.36219567592941e+00, 4.83060785179861e-03, 1.23335151497610e-02}, + {4.83060785179861e-03, 1.65289507047817e-04, -2.41490588169997e-04}, + {1.23335151497610e-02, -2.41490588169997e-04, 6.59472060689382e-03}}, + {{8.70265239309140e+00, -5.30636201431086e-04, 5.44014966585347e-03}, + {-5.30636201431086e-04, 3.11095453521008e-04, -1.86287206836035e-04}, + {5.44014966585347e-03, -1.86287206836035e-04, 6.29493388790744e-04}}, + {{4.53467851955055e+00, -3.92977536695197e-03, -2.46521420693317e-03}, + {-3.92977536695197e-03, 4.94650752632750e-05, -1.08587438501826e-05}, + {-2.46521420693317e-03, -1.08587438501826e-05, 9.28793975422261e-05}}, + {{9.26817997114275e-01, -4.03976069276753e-04, -3.56441427392165e-03}, + {-4.03976069276753e-04, 2.51976251631430e-06, 1.46914206734572e-07}, + {-3.56441427392165e-03, 1.46914206734572e-07, 8.19914567685373e-05}}, + {{7.61715986787441e+00, -1.54889041216888e-04, 2.41756280071656e-02}, + {-1.54889041216888e-04, 3.50282550461672e-07, -6.27251196972490e-06}, + {2.41756280071656e-02, -6.27251196972490e-06, 1.45061847649872e-02}}, + {{8.31193642663158e+00, -3.84070508164323e-04, -3.09750630821876e-02}, + {-3.84070508164323e-04, 3.80433432277336e-07, -1.14321142836636e-06}, + {-3.09750630821876e-02, -1.14321142836636e-06, 8.35091486289997e-04}}, + {{9.67283151270894e-01, 5.82465812445039e-05, -3.18350798617053e-03}, + {5.82465812445039e-05, 2.23762672000318e-07, -7.74196587408623e-07}, + {-3.18350798617053e-03, -7.74196587408623e-07, 3.85120938338325e-04}}, + {{8.28066236985388e+00, 5.87634508319763e-05, 6.99303090891743e-03}, + {5.87634508319763e-05, 2.93746018618058e-07, 3.40843332882272e-07}, + {6.99303090891743e-03, 3.40843332882272e-07, 1.99379171190344e-04}}, + {{6.07488998675646e+00, -1.11494526618473e-02, 5.10013111123381e-03}, + {-1.11494526618473e-02, 6.99238879921751e-04, 5.36718550370870e-05}, + {5.10013111123381e-03, 5.36718550370870e-05, 5.26909853276753e-04}}, + {{6.90492021419175e+00, 4.20639355257863e-04, -2.38612752336481e-03}, + {4.20639355257863e-04, 3.31246767338153e-06, -2.42052288150859e-08}, + {-2.38612752336481e-03, -2.42052288150859e-08, 4.46608368363412e-04}}, + {{1.31069150869715e+01, -1.73718583865670e-04, -1.97591814508578e-02}, + {-1.73718583865670e-04, 2.80451716300124e-07, 9.96570755379865e-07}, + {-1.97591814508578e-02, 9.96570755379865e-07, 2.41361900868847e-03}}, + {{4.69566344239814e+00, -2.61077567563690e-04, 5.26359000761433e-03}, + {-2.61077567563690e-04, 1.82420859823767e-06, -7.83645887541601e-07}, + {5.26359000761433e-03, -7.83645887541601e-07, 1.33586288288802e-02}}}; + +static const double kNoiseGmmMean[kNoiseGmmNumMixtures][kNoiseGmmDim] = { + {-2.01386094766163e+00, 1.69702162045397e+02, 7.41715804872181e+01}, + {-1.94684591777290e+00, 1.42398396732668e+02, 1.64186321157831e+02}, + {-2.29319297562437e+00, 3.86415425589868e+02, 2.13452215267125e+02}, + {-3.25487177070268e+00, 1.08668712553616e+03, 2.33119949467419e+02}, + {-2.13159632447467e+00, 4.83821702557717e+03, 6.86786166673740e+01}, + {-2.26171410780526e+00, 4.79420193982422e+03, 1.53222513286450e+02}, + {-3.32166740703185e+00, 4.35161135834358e+03, 1.33206448431316e+02}, + {-2.19290322814343e+00, 3.98325506609408e+03, 2.13249167359934e+02}, + {-2.02898459255404e+00, 7.37039893155007e+03, 1.12518527491926e+02}, + {-2.26150236399500e+00, 1.54896745196145e+03, 1.49717357868579e+02}, + {-2.00417668301790e+00, 3.82434760310304e+03, 1.07438913004312e+02}, + {-2.30193040814533e+00, 1.43953696546439e+03, 7.04085275122649e+01}}; + +static const double kNoiseGmmWeights[kNoiseGmmNumMixtures] = { + -1.09422832086193e+01, + -1.10847897513425e+01, + -1.36767587732187e+01, + -1.79789356118641e+01, + -1.42830169160894e+01, + -1.56500228061379e+01, + -1.83124990950113e+01, + -1.69979436177477e+01, + -1.12329424387828e+01, + -1.41311785780639e+01, + -1.47171861448585e+01, + -1.35963362781839e+01}; +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_VAD_NOISE_GMM_TABLES_H_ diff --git a/webrtc/modules/audio_processing/vad/pitch_based_vad.cc b/webrtc/modules/audio_processing/vad/pitch_based_vad.cc new file mode 100644 index 0000000000..39ec37e6ec --- /dev/null +++ b/webrtc/modules/audio_processing/vad/pitch_based_vad.cc @@ -0,0 +1,124 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/pitch_based_vad.h" + +#include <assert.h> +#include <math.h> +#include <string.h> + +#include "webrtc/modules/audio_processing/vad/vad_circular_buffer.h" +#include "webrtc/modules/audio_processing/vad/common.h" +#include "webrtc/modules/audio_processing/vad/noise_gmm_tables.h" +#include "webrtc/modules/audio_processing/vad/voice_gmm_tables.h" +#include "webrtc/modules/interface/module_common_types.h" + +namespace webrtc { + +static_assert(kNoiseGmmDim == kVoiceGmmDim, + "noise and voice gmm dimension not equal"); + +// These values should match MATLAB counterparts for unit-tests to pass. +static const int kPosteriorHistorySize = 500; // 5 sec of 10 ms frames. +static const double kInitialPriorProbability = 0.3; +static const int kTransientWidthThreshold = 7; +static const double kLowProbabilityThreshold = 0.2; + +static double LimitProbability(double p) { + const double kLimHigh = 0.99; + const double kLimLow = 0.01; + + if (p > kLimHigh) + p = kLimHigh; + else if (p < kLimLow) + p = kLimLow; + return p; +} + +PitchBasedVad::PitchBasedVad() + : p_prior_(kInitialPriorProbability), + circular_buffer_(VadCircularBuffer::Create(kPosteriorHistorySize)) { + // Setup noise GMM. + noise_gmm_.dimension = kNoiseGmmDim; + noise_gmm_.num_mixtures = kNoiseGmmNumMixtures; + noise_gmm_.weight = kNoiseGmmWeights; + noise_gmm_.mean = &kNoiseGmmMean[0][0]; + noise_gmm_.covar_inverse = &kNoiseGmmCovarInverse[0][0][0]; + + // Setup voice GMM. + voice_gmm_.dimension = kVoiceGmmDim; + voice_gmm_.num_mixtures = kVoiceGmmNumMixtures; + voice_gmm_.weight = kVoiceGmmWeights; + voice_gmm_.mean = &kVoiceGmmMean[0][0]; + voice_gmm_.covar_inverse = &kVoiceGmmCovarInverse[0][0][0]; +} + +PitchBasedVad::~PitchBasedVad() { +} + +int PitchBasedVad::VoicingProbability(const AudioFeatures& features, + double* p_combined) { + double p; + double gmm_features[3]; + double pdf_features_given_voice; + double pdf_features_given_noise; + // These limits are the same in matlab implementation 'VoicingProbGMM().' + const double kLimLowLogPitchGain = -2.0; + const double kLimHighLogPitchGain = -0.9; + const double kLimLowSpectralPeak = 200; + const double kLimHighSpectralPeak = 2000; + const double kEps = 1e-12; + for (size_t n = 0; n < features.num_frames; n++) { + gmm_features[0] = features.log_pitch_gain[n]; + gmm_features[1] = features.spectral_peak[n]; + gmm_features[2] = features.pitch_lag_hz[n]; + + pdf_features_given_voice = EvaluateGmm(gmm_features, voice_gmm_); + pdf_features_given_noise = EvaluateGmm(gmm_features, noise_gmm_); + + if (features.spectral_peak[n] < kLimLowSpectralPeak || + features.spectral_peak[n] > kLimHighSpectralPeak || + features.log_pitch_gain[n] < kLimLowLogPitchGain) { + pdf_features_given_voice = kEps * pdf_features_given_noise; + } else if (features.log_pitch_gain[n] > kLimHighLogPitchGain) { + pdf_features_given_noise = kEps * pdf_features_given_voice; + } + + p = p_prior_ * pdf_features_given_voice / + (pdf_features_given_voice * p_prior_ + + pdf_features_given_noise * (1 - p_prior_)); + + p = LimitProbability(p); + + // Combine pitch-based probability with standalone probability, before + // updating prior probabilities. + double prod_active = p * p_combined[n]; + double prod_inactive = (1 - p) * (1 - p_combined[n]); + p_combined[n] = prod_active / (prod_active + prod_inactive); + + if (UpdatePrior(p_combined[n]) < 0) + return -1; + // Limit prior probability. With a zero prior probability the posterior + // probability is always zero. + p_prior_ = LimitProbability(p_prior_); + } + return 0; +} + +int PitchBasedVad::UpdatePrior(double p) { + circular_buffer_->Insert(p); + if (circular_buffer_->RemoveTransient(kTransientWidthThreshold, + kLowProbabilityThreshold) < 0) + return -1; + p_prior_ = circular_buffer_->Mean(); + return 0; +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/pitch_based_vad.h b/webrtc/modules/audio_processing/vad/pitch_based_vad.h new file mode 100644 index 0000000000..c502184aea --- /dev/null +++ b/webrtc/modules/audio_processing/vad/pitch_based_vad.h @@ -0,0 +1,57 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_VAD_PITCH_BASED_VAD_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_VAD_PITCH_BASED_VAD_H_ + +#include "webrtc/base/scoped_ptr.h" +#include "webrtc/modules/audio_processing/vad/common.h" +#include "webrtc/modules/audio_processing/vad/gmm.h" +#include "webrtc/typedefs.h" + +namespace webrtc { + +class AudioFrame; +class VadCircularBuffer; + +// Computes the probability of the input audio frame to be active given +// the corresponding pitch-gain and lag of the frame. +class PitchBasedVad { + public: + PitchBasedVad(); + ~PitchBasedVad(); + + // Compute pitch-based voicing probability, given the features. + // features: a structure containing features required for computing voicing + // probabilities. + // + // p_combined: an array which contains the combined activity probabilities + // computed prior to the call of this function. The method, + // then, computes the voicing probabilities and combine them + // with the given values. The result are returned in |p|. + int VoicingProbability(const AudioFeatures& features, double* p_combined); + + private: + int UpdatePrior(double p); + + // TODO(turajs): maybe defining this at a higher level (maybe enum) so that + // all the code recognize it as "no-error." + static const int kNoError = 0; + + GmmParameters noise_gmm_; + GmmParameters voice_gmm_; + + double p_prior_; + + rtc::scoped_ptr<VadCircularBuffer> circular_buffer_; +}; + +} // namespace webrtc +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_VAD_PITCH_BASED_VAD_H_ diff --git a/webrtc/modules/audio_processing/vad/pitch_based_vad_unittest.cc b/webrtc/modules/audio_processing/vad/pitch_based_vad_unittest.cc new file mode 100644 index 0000000000..04ddcab5cb --- /dev/null +++ b/webrtc/modules/audio_processing/vad/pitch_based_vad_unittest.cc @@ -0,0 +1,75 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/pitch_based_vad.h" + +#include <math.h> +#include <stdio.h> + +#include <string> + +#include "testing/gtest/include/gtest/gtest.h" +#include "webrtc/test/testsupport/fileutils.h" + +namespace webrtc { + +TEST(PitchBasedVadTest, VoicingProbabilityTest) { + std::string spectral_peak_file_name = + test::ResourcePath("audio_processing/agc/agc_spectral_peak", "dat"); + FILE* spectral_peak_file = fopen(spectral_peak_file_name.c_str(), "rb"); + ASSERT_TRUE(spectral_peak_file != NULL); + + std::string pitch_gain_file_name = + test::ResourcePath("audio_processing/agc/agc_pitch_gain", "dat"); + FILE* pitch_gain_file = fopen(pitch_gain_file_name.c_str(), "rb"); + ASSERT_TRUE(pitch_gain_file != NULL); + + std::string pitch_lag_file_name = + test::ResourcePath("audio_processing/agc/agc_pitch_lag", "dat"); + FILE* pitch_lag_file = fopen(pitch_lag_file_name.c_str(), "rb"); + ASSERT_TRUE(pitch_lag_file != NULL); + + std::string voicing_prob_file_name = + test::ResourcePath("audio_processing/agc/agc_voicing_prob", "dat"); + FILE* voicing_prob_file = fopen(voicing_prob_file_name.c_str(), "rb"); + ASSERT_TRUE(voicing_prob_file != NULL); + + PitchBasedVad vad_; + + double reference_activity_probability; + + AudioFeatures audio_features; + memset(&audio_features, 0, sizeof(audio_features)); + audio_features.num_frames = 1; + while (fread(audio_features.spectral_peak, + sizeof(audio_features.spectral_peak[0]), 1, + spectral_peak_file) == 1u) { + double p; + ASSERT_EQ(1u, fread(audio_features.log_pitch_gain, + sizeof(audio_features.log_pitch_gain[0]), 1, + pitch_gain_file)); + ASSERT_EQ(1u, + fread(audio_features.pitch_lag_hz, + sizeof(audio_features.pitch_lag_hz[0]), 1, pitch_lag_file)); + ASSERT_EQ(1u, fread(&reference_activity_probability, + sizeof(reference_activity_probability), 1, + voicing_prob_file)); + + p = 0.5; // Initialize to the neutral value for combining probabilities. + EXPECT_EQ(0, vad_.VoicingProbability(audio_features, &p)); + EXPECT_NEAR(p, reference_activity_probability, 0.01); + } + + fclose(spectral_peak_file); + fclose(pitch_gain_file); + fclose(pitch_lag_file); +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/pitch_internal.cc b/webrtc/modules/audio_processing/vad/pitch_internal.cc new file mode 100644 index 0000000000..309b45acf5 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/pitch_internal.cc @@ -0,0 +1,51 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/pitch_internal.h" + +#include <cmath> + +// A 4-to-3 linear interpolation. +// The interpolation constants are derived as following: +// Input pitch parameters are updated every 7.5 ms. Within a 30-ms interval +// we are interested in pitch parameters of 0-5 ms, 10-15ms and 20-25ms. This is +// like interpolating 4-to-6 and keep the odd samples. +// The reason behind this is that LPC coefficients are computed for the first +// half of each 10ms interval. +static void PitchInterpolation(double old_val, const double* in, double* out) { + out[0] = 1. / 6. * old_val + 5. / 6. * in[0]; + out[1] = 5. / 6. * in[1] + 1. / 6. * in[2]; + out[2] = 0.5 * in[2] + 0.5 * in[3]; +} + +void GetSubframesPitchParameters(int sampling_rate_hz, + double* gains, + double* lags, + int num_in_frames, + int num_out_frames, + double* log_old_gain, + double* old_lag, + double* log_pitch_gain, + double* pitch_lag_hz) { + // Gain interpolation is in log-domain, also returned in log-domain. + for (int n = 0; n < num_in_frames; n++) + gains[n] = log(gains[n] + 1e-12); + + // Interpolate lags and gains. + PitchInterpolation(*log_old_gain, gains, log_pitch_gain); + *log_old_gain = gains[num_in_frames - 1]; + PitchInterpolation(*old_lag, lags, pitch_lag_hz); + *old_lag = lags[num_in_frames - 1]; + + // Convert pitch-lags to Hertz. + for (int n = 0; n < num_out_frames; n++) { + pitch_lag_hz[n] = (sampling_rate_hz) / (pitch_lag_hz[n]); + } +} diff --git a/webrtc/modules/audio_processing/vad/pitch_internal.h b/webrtc/modules/audio_processing/vad/pitch_internal.h new file mode 100644 index 0000000000..b25b1a82a2 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/pitch_internal.h @@ -0,0 +1,26 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_VAD_PITCH_INTERNAL_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_VAD_PITCH_INTERNAL_H_ + +// TODO(turajs): Write a description of this function. Also be consistent with +// usage of |sampling_rate_hz| vs |kSamplingFreqHz|. +void GetSubframesPitchParameters(int sampling_rate_hz, + double* gains, + double* lags, + int num_in_frames, + int num_out_frames, + double* log_old_gain, + double* old_lag, + double* log_pitch_gain, + double* pitch_lag_hz); + +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_VAD_PITCH_INTERNAL_H_ diff --git a/webrtc/modules/audio_processing/vad/pitch_internal_unittest.cc b/webrtc/modules/audio_processing/vad/pitch_internal_unittest.cc new file mode 100644 index 0000000000..8b5959d03e --- /dev/null +++ b/webrtc/modules/audio_processing/vad/pitch_internal_unittest.cc @@ -0,0 +1,50 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/pitch_internal.h" + +#include <math.h> + +#include "testing/gtest/include/gtest/gtest.h" + +TEST(PitchInternalTest, test) { + const int kSamplingRateHz = 8000; + const int kNumInputParameters = 4; + const int kNumOutputParameters = 3; + // Inputs + double log_old_gain = log(0.5); + double gains[] = {0.6, 0.2, 0.5, 0.4}; + + double old_lag = 70; + double lags[] = {90, 111, 122, 50}; + + // Expected outputs + double expected_log_pitch_gain[] = { + -0.541212549898316, -1.45672279045507, -0.80471895621705}; + double expected_log_old_gain = log(gains[kNumInputParameters - 1]); + + double expected_pitch_lag_hz[] = { + 92.3076923076923, 70.9010339734121, 93.0232558139535}; + double expected_old_lag = lags[kNumInputParameters - 1]; + + double log_pitch_gain[kNumOutputParameters]; + double pitch_lag_hz[kNumInputParameters]; + + GetSubframesPitchParameters(kSamplingRateHz, gains, lags, kNumInputParameters, + kNumOutputParameters, &log_old_gain, &old_lag, + log_pitch_gain, pitch_lag_hz); + + for (int n = 0; n < 3; n++) { + EXPECT_NEAR(pitch_lag_hz[n], expected_pitch_lag_hz[n], 1e-6); + EXPECT_NEAR(log_pitch_gain[n], expected_log_pitch_gain[n], 1e-8); + } + EXPECT_NEAR(old_lag, expected_old_lag, 1e-6); + EXPECT_NEAR(log_old_gain, expected_log_old_gain, 1e-8); +} diff --git a/webrtc/modules/audio_processing/vad/pole_zero_filter.cc b/webrtc/modules/audio_processing/vad/pole_zero_filter.cc new file mode 100644 index 0000000000..9769515c57 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/pole_zero_filter.cc @@ -0,0 +1,106 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/pole_zero_filter.h" + +#include <stdlib.h> +#include <string.h> +#include <algorithm> + +namespace webrtc { + +PoleZeroFilter* PoleZeroFilter::Create(const float* numerator_coefficients, + size_t order_numerator, + const float* denominator_coefficients, + size_t order_denominator) { + if (order_numerator > kMaxFilterOrder || + order_denominator > kMaxFilterOrder || denominator_coefficients[0] == 0 || + numerator_coefficients == NULL || denominator_coefficients == NULL) + return NULL; + return new PoleZeroFilter(numerator_coefficients, order_numerator, + denominator_coefficients, order_denominator); +} + +PoleZeroFilter::PoleZeroFilter(const float* numerator_coefficients, + size_t order_numerator, + const float* denominator_coefficients, + size_t order_denominator) + : past_input_(), + past_output_(), + numerator_coefficients_(), + denominator_coefficients_(), + order_numerator_(order_numerator), + order_denominator_(order_denominator), + highest_order_(std::max(order_denominator, order_numerator)) { + memcpy(numerator_coefficients_, numerator_coefficients, + sizeof(numerator_coefficients_[0]) * (order_numerator_ + 1)); + memcpy(denominator_coefficients_, denominator_coefficients, + sizeof(denominator_coefficients_[0]) * (order_denominator_ + 1)); + + if (denominator_coefficients_[0] != 1) { + for (size_t n = 0; n <= order_numerator_; n++) + numerator_coefficients_[n] /= denominator_coefficients_[0]; + for (size_t n = 0; n <= order_denominator_; n++) + denominator_coefficients_[n] /= denominator_coefficients_[0]; + } +} + +template <typename T> +static float FilterArPast(const T* past, size_t order, + const float* coefficients) { + float sum = 0.0f; + size_t past_index = order - 1; + for (size_t k = 1; k <= order; k++, past_index--) + sum += coefficients[k] * past[past_index]; + return sum; +} + +int PoleZeroFilter::Filter(const int16_t* in, + size_t num_input_samples, + float* output) { + if (in == NULL || output == NULL) + return -1; + // This is the typical case, just a memcpy. + const size_t k = std::min(num_input_samples, highest_order_); + size_t n; + for (n = 0; n < k; n++) { + output[n] = in[n] * numerator_coefficients_[0]; + output[n] += FilterArPast(&past_input_[n], order_numerator_, + numerator_coefficients_); + output[n] -= FilterArPast(&past_output_[n], order_denominator_, + denominator_coefficients_); + + past_input_[n + order_numerator_] = in[n]; + past_output_[n + order_denominator_] = output[n]; + } + if (highest_order_ < num_input_samples) { + for (size_t m = 0; n < num_input_samples; n++, m++) { + output[n] = in[n] * numerator_coefficients_[0]; + output[n] += + FilterArPast(&in[m], order_numerator_, numerator_coefficients_); + output[n] -= FilterArPast(&output[m], order_denominator_, + denominator_coefficients_); + } + // Record into the past signal. + memcpy(past_input_, &in[num_input_samples - order_numerator_], + sizeof(in[0]) * order_numerator_); + memcpy(past_output_, &output[num_input_samples - order_denominator_], + sizeof(output[0]) * order_denominator_); + } else { + // Odd case that the length of the input is shorter that filter order. + memmove(past_input_, &past_input_[num_input_samples], + order_numerator_ * sizeof(past_input_[0])); + memmove(past_output_, &past_output_[num_input_samples], + order_denominator_ * sizeof(past_output_[0])); + } + return 0; +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/pole_zero_filter.h b/webrtc/modules/audio_processing/vad/pole_zero_filter.h new file mode 100644 index 0000000000..bd13050a5c --- /dev/null +++ b/webrtc/modules/audio_processing/vad/pole_zero_filter.h @@ -0,0 +1,52 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_VAD_POLE_ZERO_FILTER_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_VAD_POLE_ZERO_FILTER_H_ + +#include <cstddef> + +#include "webrtc/typedefs.h" + +namespace webrtc { + +class PoleZeroFilter { + public: + ~PoleZeroFilter() {} + + static PoleZeroFilter* Create(const float* numerator_coefficients, + size_t order_numerator, + const float* denominator_coefficients, + size_t order_denominator); + + int Filter(const int16_t* in, size_t num_input_samples, float* output); + + private: + PoleZeroFilter(const float* numerator_coefficients, + size_t order_numerator, + const float* denominator_coefficients, + size_t order_denominator); + + static const int kMaxFilterOrder = 24; + + int16_t past_input_[kMaxFilterOrder * 2]; + float past_output_[kMaxFilterOrder * 2]; + + float numerator_coefficients_[kMaxFilterOrder + 1]; + float denominator_coefficients_[kMaxFilterOrder + 1]; + + size_t order_numerator_; + size_t order_denominator_; + size_t highest_order_; +}; + +} // namespace webrtc + +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_VAD_POLE_ZERO_FILTER_H_ diff --git a/webrtc/modules/audio_processing/vad/pole_zero_filter_unittest.cc b/webrtc/modules/audio_processing/vad/pole_zero_filter_unittest.cc new file mode 100644 index 0000000000..492c3f0c94 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/pole_zero_filter_unittest.cc @@ -0,0 +1,102 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/pole_zero_filter.h" + +#include <math.h> +#include <stdio.h> + +#include "testing/gtest/include/gtest/gtest.h" +#include "webrtc/base/scoped_ptr.h" +#include "webrtc/modules/audio_processing/vad/vad_audio_proc_internal.h" +#include "webrtc/test/testsupport/fileutils.h" + +namespace webrtc { + +static const int kInputSamples = 50; + +static const int16_t kInput[kInputSamples] = { + -2136, -7116, 10715, 2464, 3164, 8139, 11393, 24013, -32117, -5544, + -27740, 10181, 14190, -24055, -15912, 17393, 6359, -9950, -13894, 32432, + -23944, 3437, -8381, 19768, 3087, -19795, -5920, 13310, 1407, 3876, + 4059, 3524, -23130, 19121, -27900, -24840, 4089, 21422, -3625, 3015, + -11236, 28856, 13424, 6571, -19761, -6361, 15821, -9469, 29727, 32229}; + +static const float kReferenceOutput[kInputSamples] = { + -2082.230472f, -6878.572941f, 10697.090871f, 2358.373952f, + 2973.936512f, 7738.580650f, 10690.803213f, 22687.091576f, + -32676.684717f, -5879.621684f, -27359.297432f, 10368.735888f, + 13994.584604f, -23676.126249f, -15078.250390f, 17818.253338f, + 6577.743123f, -9498.369315f, -13073.651079f, 32460.026588f, + -23391.849347f, 3953.805667f, -7667.761363f, 19995.153447f, + 3185.575477f, -19207.365160f, -5143.103201f, 13756.317237f, + 1779.654794f, 4142.269755f, 4209.475034f, 3572.991789f, + -22509.089546f, 19307.878964f, -27060.439759f, -23319.042810f, + 5547.685267f, 22312.718676f, -2707.309027f, 3852.358490f, + -10135.510093f, 29241.509970f, 13394.397233f, 6340.721417f, + -19510.207905f, -5908.442086f, 15882.301634f, -9211.335255f, + 29253.056735f, 30874.443046f}; + +class PoleZeroFilterTest : public ::testing::Test { + protected: + PoleZeroFilterTest() + : my_filter_(PoleZeroFilter::Create(kCoeffNumerator, + kFilterOrder, + kCoeffDenominator, + kFilterOrder)) {} + + ~PoleZeroFilterTest() {} + + void FilterSubframes(int num_subframes); + + private: + void TestClean(); + rtc::scoped_ptr<PoleZeroFilter> my_filter_; +}; + +void PoleZeroFilterTest::FilterSubframes(int num_subframes) { + float output[kInputSamples]; + const int num_subframe_samples = kInputSamples / num_subframes; + EXPECT_EQ(num_subframe_samples * num_subframes, kInputSamples); + + for (int n = 0; n < num_subframes; n++) { + my_filter_->Filter(&kInput[n * num_subframe_samples], num_subframe_samples, + &output[n * num_subframe_samples]); + } + for (int n = 0; n < kInputSamples; n++) { + EXPECT_NEAR(output[n], kReferenceOutput[n], 1); + } +} + +TEST_F(PoleZeroFilterTest, OneSubframe) { + FilterSubframes(1); +} + +TEST_F(PoleZeroFilterTest, TwoSubframes) { + FilterSubframes(2); +} + +TEST_F(PoleZeroFilterTest, FiveSubframes) { + FilterSubframes(5); +} + +TEST_F(PoleZeroFilterTest, TenSubframes) { + FilterSubframes(10); +} + +TEST_F(PoleZeroFilterTest, TwentyFiveSubframes) { + FilterSubframes(25); +} + +TEST_F(PoleZeroFilterTest, FiftySubframes) { + FilterSubframes(50); +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/standalone_vad.cc b/webrtc/modules/audio_processing/vad/standalone_vad.cc new file mode 100644 index 0000000000..468b8ff3f0 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/standalone_vad.cc @@ -0,0 +1,93 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/standalone_vad.h" + +#include <assert.h> + +#include "webrtc/modules/interface/module_common_types.h" +#include "webrtc/modules/utility/interface/audio_frame_operations.h" +#include "webrtc/typedefs.h" + +namespace webrtc { + +static const int kDefaultStandaloneVadMode = 3; + +StandaloneVad::StandaloneVad(VadInst* vad) + : vad_(vad), buffer_(), index_(0), mode_(kDefaultStandaloneVadMode) { +} + +StandaloneVad::~StandaloneVad() { + WebRtcVad_Free(vad_); +} + +StandaloneVad* StandaloneVad::Create() { + VadInst* vad = WebRtcVad_Create(); + if (!vad) + return nullptr; + + int err = WebRtcVad_Init(vad); + err |= WebRtcVad_set_mode(vad, kDefaultStandaloneVadMode); + if (err != 0) { + WebRtcVad_Free(vad); + return nullptr; + } + return new StandaloneVad(vad); +} + +int StandaloneVad::AddAudio(const int16_t* data, size_t length) { + if (length != kLength10Ms) + return -1; + + if (index_ + length > kLength10Ms * kMaxNum10msFrames) + // Reset the buffer if it's full. + // TODO(ajm): Instead, consider just processing every 10 ms frame. Then we + // can forgo the buffering. + index_ = 0; + + memcpy(&buffer_[index_], data, sizeof(int16_t) * length); + index_ += length; + return 0; +} + +int StandaloneVad::GetActivity(double* p, size_t length_p) { + if (index_ == 0) + return -1; + + const size_t num_frames = index_ / kLength10Ms; + if (num_frames > length_p) + return -1; + assert(WebRtcVad_ValidRateAndFrameLength(kSampleRateHz, index_) == 0); + + int activity = WebRtcVad_Process(vad_, kSampleRateHz, buffer_, index_); + if (activity < 0) + return -1; + else if (activity == 0) + p[0] = 0.01; // Arbitrary but small and non-zero. + else + p[0] = 0.5; // 0.5 is neutral values when combinned by other probabilities. + for (size_t n = 1; n < num_frames; n++) + p[n] = p[0]; + // Reset the buffer to start from the beginning. + index_ = 0; + return activity; +} + +int StandaloneVad::set_mode(int mode) { + if (mode < 0 || mode > 3) + return -1; + if (WebRtcVad_set_mode(vad_, mode) != 0) + return -1; + + mode_ = mode; + return 0; +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/standalone_vad.h b/webrtc/modules/audio_processing/vad/standalone_vad.h new file mode 100644 index 0000000000..6a25424dab --- /dev/null +++ b/webrtc/modules/audio_processing/vad/standalone_vad.h @@ -0,0 +1,70 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AGC_STANDALONE_VAD_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_AGC_STANDALONE_VAD_H_ + +#include "webrtc/base/scoped_ptr.h" +#include "webrtc/modules/audio_processing/vad/common.h" +#include "webrtc/common_audio/vad/include/webrtc_vad.h" +#include "webrtc/typedefs.h" + +namespace webrtc { + +class AudioFrame; + +class StandaloneVad { + public: + static StandaloneVad* Create(); + ~StandaloneVad(); + + // Outputs + // p: a buffer where probabilities are written to. + // length_p: number of elements of |p|. + // + // return value: + // -1: if no audio is stored or VAD returns error. + // 0: in success. + // In case of error the content of |activity| is unchanged. + // + // Note that due to a high false-positive (VAD decision is active while the + // processed audio is just background noise) rate, stand-alone VAD is used as + // a one-sided indicator. The activity probability is 0.5 if the frame is + // classified as active, and the probability is 0.01 if the audio is + // classified as passive. In this way, when probabilities are combined, the + // effect of the stand-alone VAD is neutral if the input is classified as + // active. + int GetActivity(double* p, size_t length_p); + + // Expecting 10 ms of 16 kHz audio to be pushed in. + int AddAudio(const int16_t* data, size_t length); + + // Set aggressiveness of VAD, 0 is the least aggressive and 3 is the most + // aggressive mode. Returns -1 if the input is less than 0 or larger than 3, + // otherwise 0 is returned. + int set_mode(int mode); + // Get the agressiveness of the current VAD. + int mode() const { return mode_; } + + private: + explicit StandaloneVad(VadInst* vad); + + static const size_t kMaxNum10msFrames = 3; + + // TODO(turajs): Is there a way to use scoped-pointer here? + VadInst* vad_; + int16_t buffer_[kMaxNum10msFrames * kLength10Ms]; + size_t index_; + int mode_; +}; + +} // namespace webrtc + +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_AGC_STANDALONE_VAD_H_ diff --git a/webrtc/modules/audio_processing/vad/standalone_vad_unittest.cc b/webrtc/modules/audio_processing/vad/standalone_vad_unittest.cc new file mode 100644 index 0000000000..942008e733 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/standalone_vad_unittest.cc @@ -0,0 +1,104 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/standalone_vad.h" + +#include <string.h> + +#include "testing/gtest/include/gtest/gtest.h" +#include "webrtc/base/scoped_ptr.h" +#include "webrtc/modules/interface/module_common_types.h" +#include "webrtc/test/testsupport/fileutils.h" +#include "webrtc/test/testsupport/gtest_disable.h" + +namespace webrtc { + +TEST(StandaloneVadTest, Api) { + rtc::scoped_ptr<StandaloneVad> vad(StandaloneVad::Create()); + int16_t data[kLength10Ms] = {0}; + + // Valid frame length (for 32 kHz rate), but not what the VAD is expecting. + EXPECT_EQ(-1, vad->AddAudio(data, 320)); + + const size_t kMaxNumFrames = 3; + double p[kMaxNumFrames]; + for (size_t n = 0; n < kMaxNumFrames; n++) + EXPECT_EQ(0, vad->AddAudio(data, kLength10Ms)); + + // Pretend |p| is shorter that it should be. + EXPECT_EQ(-1, vad->GetActivity(p, kMaxNumFrames - 1)); + + EXPECT_EQ(0, vad->GetActivity(p, kMaxNumFrames)); + + // Ask for activity when buffer is empty. + EXPECT_EQ(-1, vad->GetActivity(p, kMaxNumFrames)); + + // Should reset and result in one buffer. + for (size_t n = 0; n < kMaxNumFrames + 1; n++) + EXPECT_EQ(0, vad->AddAudio(data, kLength10Ms)); + EXPECT_EQ(0, vad->GetActivity(p, 1)); + + // Wrong modes + EXPECT_EQ(-1, vad->set_mode(-1)); + EXPECT_EQ(-1, vad->set_mode(4)); + + // Valid mode. + const int kMode = 2; + EXPECT_EQ(0, vad->set_mode(kMode)); + EXPECT_EQ(kMode, vad->mode()); +} + +TEST(StandaloneVadTest, DISABLED_ON_IOS(ActivityDetection)) { + rtc::scoped_ptr<StandaloneVad> vad(StandaloneVad::Create()); + const size_t kDataLength = kLength10Ms; + int16_t data[kDataLength] = {0}; + + FILE* pcm_file = + fopen(test::ResourcePath("audio_processing/agc/agc_audio", "pcm").c_str(), + "rb"); + ASSERT_TRUE(pcm_file != NULL); + + FILE* reference_file = fopen( + test::ResourcePath("audio_processing/agc/agc_vad", "dat").c_str(), "rb"); + ASSERT_TRUE(reference_file != NULL); + + // Reference activities are prepared with 0 aggressiveness. + ASSERT_EQ(0, vad->set_mode(0)); + + // Stand-alone VAD can operate on 1, 2 or 3 frames of length 10 ms. The + // reference file is created for 30 ms frame. + const int kNumVadFramesToProcess = 3; + int num_frames = 0; + while (fread(data, sizeof(int16_t), kDataLength, pcm_file) == kDataLength) { + vad->AddAudio(data, kDataLength); + num_frames++; + if (num_frames == kNumVadFramesToProcess) { + num_frames = 0; + int referece_activity; + double p[kNumVadFramesToProcess]; + EXPECT_EQ(1u, fread(&referece_activity, sizeof(referece_activity), 1, + reference_file)); + int activity = vad->GetActivity(p, kNumVadFramesToProcess); + EXPECT_EQ(referece_activity, activity); + if (activity != 0) { + // When active, probabilities are set to 0.5. + for (int n = 0; n < kNumVadFramesToProcess; n++) + EXPECT_EQ(0.5, p[n]); + } else { + // When inactive, probabilities are set to 0.01. + for (int n = 0; n < kNumVadFramesToProcess; n++) + EXPECT_EQ(0.01, p[n]); + } + } + } + fclose(reference_file); + fclose(pcm_file); +} +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/vad_audio_proc.cc b/webrtc/modules/audio_processing/vad/vad_audio_proc.cc new file mode 100644 index 0000000000..8535d1ff57 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/vad_audio_proc.cc @@ -0,0 +1,275 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/vad_audio_proc.h" + +#include <math.h> +#include <stdio.h> + +#include "webrtc/common_audio/fft4g.h" +#include "webrtc/modules/audio_processing/vad/vad_audio_proc_internal.h" +#include "webrtc/modules/audio_processing/vad/pitch_internal.h" +#include "webrtc/modules/audio_processing/vad/pole_zero_filter.h" +extern "C" { +#include "webrtc/modules/audio_coding/codecs/isac/main/source/codec.h" +#include "webrtc/modules/audio_coding/codecs/isac/main/source/lpc_analysis.h" +#include "webrtc/modules/audio_coding/codecs/isac/main/source/pitch_estimator.h" +#include "webrtc/modules/audio_coding/codecs/isac/main/source/structs.h" +} +#include "webrtc/modules/interface/module_common_types.h" + +namespace webrtc { + +// The following structures are declared anonymous in iSAC's structs.h. To +// forward declare them, we use this derived class trick. +struct VadAudioProc::PitchAnalysisStruct : public ::PitchAnalysisStruct {}; +struct VadAudioProc::PreFiltBankstr : public ::PreFiltBankstr {}; + +static const float kFrequencyResolution = + kSampleRateHz / static_cast<float>(VadAudioProc::kDftSize); +static const int kSilenceRms = 5; + +// TODO(turajs): Make a Create or Init for VadAudioProc. +VadAudioProc::VadAudioProc() + : audio_buffer_(), + num_buffer_samples_(kNumPastSignalSamples), + log_old_gain_(-2), + old_lag_(50), // Arbitrary but valid as pitch-lag (in samples). + pitch_analysis_handle_(new PitchAnalysisStruct), + pre_filter_handle_(new PreFiltBankstr), + high_pass_filter_(PoleZeroFilter::Create(kCoeffNumerator, + kFilterOrder, + kCoeffDenominator, + kFilterOrder)) { + static_assert(kNumPastSignalSamples + kNumSubframeSamples == + sizeof(kLpcAnalWin) / sizeof(kLpcAnalWin[0]), + "lpc analysis window incorrect size"); + static_assert(kLpcOrder + 1 == sizeof(kCorrWeight) / sizeof(kCorrWeight[0]), + "correlation weight incorrect size"); + + // TODO(turajs): Are we doing too much in the constructor? + float data[kDftSize]; + // Make FFT to initialize. + ip_[0] = 0; + WebRtc_rdft(kDftSize, 1, data, ip_, w_fft_); + // TODO(turajs): Need to initialize high-pass filter. + + // Initialize iSAC components. + WebRtcIsac_InitPreFilterbank(pre_filter_handle_.get()); + WebRtcIsac_InitPitchAnalysis(pitch_analysis_handle_.get()); +} + +VadAudioProc::~VadAudioProc() { +} + +void VadAudioProc::ResetBuffer() { + memcpy(audio_buffer_, &audio_buffer_[kNumSamplesToProcess], + sizeof(audio_buffer_[0]) * kNumPastSignalSamples); + num_buffer_samples_ = kNumPastSignalSamples; +} + +int VadAudioProc::ExtractFeatures(const int16_t* frame, + size_t length, + AudioFeatures* features) { + features->num_frames = 0; + if (length != kNumSubframeSamples) { + return -1; + } + + // High-pass filter to remove the DC component and very low frequency content. + // We have experienced that this high-pass filtering improves voice/non-voiced + // classification. + if (high_pass_filter_->Filter(frame, kNumSubframeSamples, + &audio_buffer_[num_buffer_samples_]) != 0) { + return -1; + } + + num_buffer_samples_ += kNumSubframeSamples; + if (num_buffer_samples_ < kBufferLength) { + return 0; + } + assert(num_buffer_samples_ == kBufferLength); + features->num_frames = kNum10msSubframes; + features->silence = false; + + Rms(features->rms, kMaxNumFrames); + for (size_t i = 0; i < kNum10msSubframes; ++i) { + if (features->rms[i] < kSilenceRms) { + // PitchAnalysis can cause NaNs in the pitch gain if it's fed silence. + // Bail out here instead. + features->silence = true; + ResetBuffer(); + return 0; + } + } + + PitchAnalysis(features->log_pitch_gain, features->pitch_lag_hz, + kMaxNumFrames); + FindFirstSpectralPeaks(features->spectral_peak, kMaxNumFrames); + ResetBuffer(); + return 0; +} + +// Computes |kLpcOrder + 1| correlation coefficients. +void VadAudioProc::SubframeCorrelation(double* corr, + size_t length_corr, + size_t subframe_index) { + assert(length_corr >= kLpcOrder + 1); + double windowed_audio[kNumSubframeSamples + kNumPastSignalSamples]; + size_t buffer_index = subframe_index * kNumSubframeSamples; + + for (size_t n = 0; n < kNumSubframeSamples + kNumPastSignalSamples; n++) + windowed_audio[n] = audio_buffer_[buffer_index++] * kLpcAnalWin[n]; + + WebRtcIsac_AutoCorr(corr, windowed_audio, + kNumSubframeSamples + kNumPastSignalSamples, kLpcOrder); +} + +// Compute |kNum10msSubframes| sets of LPC coefficients, one per 10 ms input. +// The analysis window is 15 ms long and it is centered on the first half of +// each 10ms sub-frame. This is equivalent to computing LPC coefficients for the +// first half of each 10 ms subframe. +void VadAudioProc::GetLpcPolynomials(double* lpc, size_t length_lpc) { + assert(length_lpc >= kNum10msSubframes * (kLpcOrder + 1)); + double corr[kLpcOrder + 1]; + double reflec_coeff[kLpcOrder]; + for (size_t i = 0, offset_lpc = 0; i < kNum10msSubframes; + i++, offset_lpc += kLpcOrder + 1) { + SubframeCorrelation(corr, kLpcOrder + 1, i); + corr[0] *= 1.0001; + // This makes Lev-Durb a bit more stable. + for (size_t k = 0; k < kLpcOrder + 1; k++) { + corr[k] *= kCorrWeight[k]; + } + WebRtcIsac_LevDurb(&lpc[offset_lpc], reflec_coeff, corr, kLpcOrder); + } +} + +// Fit a second order curve to these 3 points and find the location of the +// extremum. The points are inverted before curve fitting. +static float QuadraticInterpolation(float prev_val, + float curr_val, + float next_val) { + // Doing the interpolation in |1 / A(z)|^2. + float fractional_index = 0; + next_val = 1.0f / next_val; + prev_val = 1.0f / prev_val; + curr_val = 1.0f / curr_val; + + fractional_index = + -(next_val - prev_val) * 0.5f / (next_val + prev_val - 2.f * curr_val); + assert(fabs(fractional_index) < 1); + return fractional_index; +} + +// 1 / A(z), where A(z) is defined by |lpc| is a model of the spectral envelope +// of the input signal. The local maximum of the spectral envelope corresponds +// with the local minimum of A(z). It saves complexity, as we save one +// inversion. Furthermore, we find the first local maximum of magnitude squared, +// to save on one square root. +void VadAudioProc::FindFirstSpectralPeaks(double* f_peak, + size_t length_f_peak) { + assert(length_f_peak >= kNum10msSubframes); + double lpc[kNum10msSubframes * (kLpcOrder + 1)]; + // For all sub-frames. + GetLpcPolynomials(lpc, kNum10msSubframes * (kLpcOrder + 1)); + + const size_t kNumDftCoefficients = kDftSize / 2 + 1; + float data[kDftSize]; + + for (size_t i = 0; i < kNum10msSubframes; i++) { + // Convert to float with zero pad. + memset(data, 0, sizeof(data)); + for (size_t n = 0; n < kLpcOrder + 1; n++) { + data[n] = static_cast<float>(lpc[i * (kLpcOrder + 1) + n]); + } + // Transform to frequency domain. + WebRtc_rdft(kDftSize, 1, data, ip_, w_fft_); + + size_t index_peak = 0; + float prev_magn_sqr = data[0] * data[0]; + float curr_magn_sqr = data[2] * data[2] + data[3] * data[3]; + float next_magn_sqr; + bool found_peak = false; + for (size_t n = 2; n < kNumDftCoefficients - 1; n++) { + next_magn_sqr = + data[2 * n] * data[2 * n] + data[2 * n + 1] * data[2 * n + 1]; + if (curr_magn_sqr < prev_magn_sqr && curr_magn_sqr < next_magn_sqr) { + found_peak = true; + index_peak = n - 1; + break; + } + prev_magn_sqr = curr_magn_sqr; + curr_magn_sqr = next_magn_sqr; + } + float fractional_index = 0; + if (!found_peak) { + // Checking if |kNumDftCoefficients - 1| is the local minimum. + next_magn_sqr = data[1] * data[1]; + if (curr_magn_sqr < prev_magn_sqr && curr_magn_sqr < next_magn_sqr) { + index_peak = kNumDftCoefficients - 1; + } + } else { + // A peak is found, do a simple quadratic interpolation to get a more + // accurate estimate of the peak location. + fractional_index = + QuadraticInterpolation(prev_magn_sqr, curr_magn_sqr, next_magn_sqr); + } + f_peak[i] = (index_peak + fractional_index) * kFrequencyResolution; + } +} + +// Using iSAC functions to estimate pitch gains & lags. +void VadAudioProc::PitchAnalysis(double* log_pitch_gains, + double* pitch_lags_hz, + size_t length) { + // TODO(turajs): This can be "imported" from iSAC & and the next two + // constants. + assert(length >= kNum10msSubframes); + const int kNumPitchSubframes = 4; + double gains[kNumPitchSubframes]; + double lags[kNumPitchSubframes]; + + const int kNumSubbandFrameSamples = 240; + const int kNumLookaheadSamples = 24; + + float lower[kNumSubbandFrameSamples]; + float upper[kNumSubbandFrameSamples]; + double lower_lookahead[kNumSubbandFrameSamples]; + double upper_lookahead[kNumSubbandFrameSamples]; + double lower_lookahead_pre_filter[kNumSubbandFrameSamples + + kNumLookaheadSamples]; + + // Split signal to lower and upper bands + WebRtcIsac_SplitAndFilterFloat(&audio_buffer_[kNumPastSignalSamples], lower, + upper, lower_lookahead, upper_lookahead, + pre_filter_handle_.get()); + WebRtcIsac_PitchAnalysis(lower_lookahead, lower_lookahead_pre_filter, + pitch_analysis_handle_.get(), lags, gains); + + // Lags are computed on lower-band signal with sampling rate half of the + // input signal. + GetSubframesPitchParameters( + kSampleRateHz / 2, gains, lags, kNumPitchSubframes, kNum10msSubframes, + &log_old_gain_, &old_lag_, log_pitch_gains, pitch_lags_hz); +} + +void VadAudioProc::Rms(double* rms, size_t length_rms) { + assert(length_rms >= kNum10msSubframes); + size_t offset = kNumPastSignalSamples; + for (size_t i = 0; i < kNum10msSubframes; i++) { + rms[i] = 0; + for (size_t n = 0; n < kNumSubframeSamples; n++, offset++) + rms[i] += audio_buffer_[offset] * audio_buffer_[offset]; + rms[i] = sqrt(rms[i] / kNumSubframeSamples); + } +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/vad_audio_proc.h b/webrtc/modules/audio_processing/vad/vad_audio_proc.h new file mode 100644 index 0000000000..85500aed84 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/vad_audio_proc.h @@ -0,0 +1,89 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VAD_AUDIO_PROC_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VAD_AUDIO_PROC_H_ + +#include "webrtc/base/scoped_ptr.h" +#include "webrtc/modules/audio_processing/vad/common.h" +#include "webrtc/typedefs.h" + +namespace webrtc { + +class AudioFrame; +class PoleZeroFilter; + +class VadAudioProc { + public: + // Forward declare iSAC structs. + struct PitchAnalysisStruct; + struct PreFiltBankstr; + + VadAudioProc(); + ~VadAudioProc(); + + int ExtractFeatures(const int16_t* audio_frame, + size_t length, + AudioFeatures* audio_features); + + static const size_t kDftSize = 512; + + private: + void PitchAnalysis(double* pitch_gains, double* pitch_lags_hz, size_t length); + void SubframeCorrelation(double* corr, + size_t length_corr, + size_t subframe_index); + void GetLpcPolynomials(double* lpc, size_t length_lpc); + void FindFirstSpectralPeaks(double* f_peak, size_t length_f_peak); + void Rms(double* rms, size_t length_rms); + void ResetBuffer(); + + // To compute spectral peak we perform LPC analysis to get spectral envelope. + // For every 30 ms we compute 3 spectral peak there for 3 LPC analysis. + // LPC is computed over 15 ms of windowed audio. For every 10 ms sub-frame + // we need 5 ms of past signal to create the input of LPC analysis. + static const size_t kNumPastSignalSamples = + static_cast<size_t>(kSampleRateHz / 200); + + // TODO(turajs): maybe defining this at a higher level (maybe enum) so that + // all the code recognize it as "no-error." + static const int kNoError = 0; + + static const size_t kNum10msSubframes = 3; + static const size_t kNumSubframeSamples = + static_cast<size_t>(kSampleRateHz / 100); + static const size_t kNumSamplesToProcess = + kNum10msSubframes * + kNumSubframeSamples; // Samples in 30 ms @ given sampling rate. + static const size_t kBufferLength = + kNumPastSignalSamples + kNumSamplesToProcess; + static const size_t kIpLength = kDftSize >> 1; + static const size_t kWLength = kDftSize >> 1; + + static const size_t kLpcOrder = 16; + + size_t ip_[kIpLength]; + float w_fft_[kWLength]; + + // A buffer of 5 ms (past audio) + 30 ms (one iSAC frame ). + float audio_buffer_[kBufferLength]; + size_t num_buffer_samples_; + + double log_old_gain_; + double old_lag_; + + rtc::scoped_ptr<PitchAnalysisStruct> pitch_analysis_handle_; + rtc::scoped_ptr<PreFiltBankstr> pre_filter_handle_; + rtc::scoped_ptr<PoleZeroFilter> high_pass_filter_; +}; + +} // namespace webrtc + +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VAD_AUDIO_PROC_H_ diff --git a/webrtc/modules/audio_processing/vad/vad_audio_proc_internal.h b/webrtc/modules/audio_processing/vad/vad_audio_proc_internal.h new file mode 100644 index 0000000000..45586b9be6 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/vad_audio_proc_internal.h @@ -0,0 +1,94 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VAD_AUDIO_PROC_INTERNAL_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VAD_AUDIO_PROC_INTERNAL_H_ + +namespace webrtc { + +// These values should match MATLAB counterparts for unit-tests to pass. +static const double kCorrWeight[] = {1.000000, + 0.985000, + 0.970225, + 0.955672, + 0.941337, + 0.927217, + 0.913308, + 0.899609, + 0.886115, + 0.872823, + 0.859730, + 0.846834, + 0.834132, + 0.821620, + 0.809296, + 0.797156, + 0.785199}; + +static const double kLpcAnalWin[] = { + 0.00000000, 0.01314436, 0.02628645, 0.03942400, 0.05255473, 0.06567639, + 0.07878670, 0.09188339, 0.10496421, 0.11802689, 0.13106918, 0.14408883, + 0.15708358, 0.17005118, 0.18298941, 0.19589602, 0.20876878, 0.22160547, + 0.23440387, 0.24716177, 0.25987696, 0.27254725, 0.28517045, 0.29774438, + 0.31026687, 0.32273574, 0.33514885, 0.34750406, 0.35979922, 0.37203222, + 0.38420093, 0.39630327, 0.40833713, 0.42030043, 0.43219112, 0.44400713, + 0.45574642, 0.46740697, 0.47898676, 0.49048379, 0.50189608, 0.51322164, + 0.52445853, 0.53560481, 0.54665854, 0.55761782, 0.56848075, 0.57924546, + 0.58991008, 0.60047278, 0.61093173, 0.62128512, 0.63153117, 0.64166810, + 0.65169416, 0.66160761, 0.67140676, 0.68108990, 0.69065536, 0.70010148, + 0.70942664, 0.71862923, 0.72770765, 0.73666033, 0.74548573, 0.75418233, + 0.76274862, 0.77118312, 0.77948437, 0.78765094, 0.79568142, 0.80357442, + 0.81132858, 0.81894256, 0.82641504, 0.83374472, 0.84093036, 0.84797069, + 0.85486451, 0.86161063, 0.86820787, 0.87465511, 0.88095122, 0.88709512, + 0.89308574, 0.89892206, 0.90460306, 0.91012776, 0.91549520, 0.92070447, + 0.92575465, 0.93064488, 0.93537432, 0.93994213, 0.94434755, 0.94858979, + 0.95266814, 0.95658189, 0.96033035, 0.96391289, 0.96732888, 0.97057773, + 0.97365889, 0.97657181, 0.97931600, 0.98189099, 0.98429632, 0.98653158, + 0.98859639, 0.99049038, 0.99221324, 0.99376466, 0.99514438, 0.99635215, + 0.99738778, 0.99825107, 0.99894188, 0.99946010, 0.99980562, 0.99997840, + 0.99997840, 0.99980562, 0.99946010, 0.99894188, 0.99825107, 0.99738778, + 0.99635215, 0.99514438, 0.99376466, 0.99221324, 0.99049038, 0.98859639, + 0.98653158, 0.98429632, 0.98189099, 0.97931600, 0.97657181, 0.97365889, + 0.97057773, 0.96732888, 0.96391289, 0.96033035, 0.95658189, 0.95266814, + 0.94858979, 0.94434755, 0.93994213, 0.93537432, 0.93064488, 0.92575465, + 0.92070447, 0.91549520, 0.91012776, 0.90460306, 0.89892206, 0.89308574, + 0.88709512, 0.88095122, 0.87465511, 0.86820787, 0.86161063, 0.85486451, + 0.84797069, 0.84093036, 0.83374472, 0.82641504, 0.81894256, 0.81132858, + 0.80357442, 0.79568142, 0.78765094, 0.77948437, 0.77118312, 0.76274862, + 0.75418233, 0.74548573, 0.73666033, 0.72770765, 0.71862923, 0.70942664, + 0.70010148, 0.69065536, 0.68108990, 0.67140676, 0.66160761, 0.65169416, + 0.64166810, 0.63153117, 0.62128512, 0.61093173, 0.60047278, 0.58991008, + 0.57924546, 0.56848075, 0.55761782, 0.54665854, 0.53560481, 0.52445853, + 0.51322164, 0.50189608, 0.49048379, 0.47898676, 0.46740697, 0.45574642, + 0.44400713, 0.43219112, 0.42030043, 0.40833713, 0.39630327, 0.38420093, + 0.37203222, 0.35979922, 0.34750406, 0.33514885, 0.32273574, 0.31026687, + 0.29774438, 0.28517045, 0.27254725, 0.25987696, 0.24716177, 0.23440387, + 0.22160547, 0.20876878, 0.19589602, 0.18298941, 0.17005118, 0.15708358, + 0.14408883, 0.13106918, 0.11802689, 0.10496421, 0.09188339, 0.07878670, + 0.06567639, 0.05255473, 0.03942400, 0.02628645, 0.01314436, 0.00000000}; + +static const size_t kFilterOrder = 2; +static const float kCoeffNumerator[kFilterOrder + 1] = {0.974827f, + -1.949650f, + 0.974827f}; +static const float kCoeffDenominator[kFilterOrder + 1] = {1.0f, + -1.971999f, + 0.972457f}; + +static_assert(kFilterOrder + 1 == + sizeof(kCoeffNumerator) / sizeof(kCoeffNumerator[0]), + "numerator coefficients incorrect size"); +static_assert(kFilterOrder + 1 == + sizeof(kCoeffDenominator) / sizeof(kCoeffDenominator[0]), + "denominator coefficients incorrect size"); + +} // namespace webrtc + +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VAD_AUDIO_PROCESSING_H_ diff --git a/webrtc/modules/audio_processing/vad/vad_audio_proc_unittest.cc b/webrtc/modules/audio_processing/vad/vad_audio_proc_unittest.cc new file mode 100644 index 0000000000..f509af476f --- /dev/null +++ b/webrtc/modules/audio_processing/vad/vad_audio_proc_unittest.cc @@ -0,0 +1,63 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +// We don't test the value of pitch gain and lags as they are created by iSAC +// routines. However, interpolation of pitch-gain and lags is in a separate +// class and has its own unit-test. + +#include "webrtc/modules/audio_processing/vad/vad_audio_proc.h" + +#include <math.h> +#include <stdio.h> + +#include <string> + +#include "testing/gtest/include/gtest/gtest.h" +#include "webrtc/modules/audio_processing/vad/common.h" +#include "webrtc/modules/interface/module_common_types.h" +#include "webrtc/test/testsupport/fileutils.h" + +namespace webrtc { + +TEST(AudioProcessingTest, DISABLED_ComputingFirstSpectralPeak) { + VadAudioProc audioproc; + + std::string peak_file_name = + test::ResourcePath("audio_processing/agc/agc_spectral_peak", "dat"); + FILE* peak_file = fopen(peak_file_name.c_str(), "rb"); + ASSERT_TRUE(peak_file != NULL); + + std::string pcm_file_name = + test::ResourcePath("audio_processing/agc/agc_audio", "pcm"); + FILE* pcm_file = fopen(pcm_file_name.c_str(), "rb"); + ASSERT_TRUE(pcm_file != NULL); + + // Read 10 ms audio in each iteration. + const size_t kDataLength = kLength10Ms; + int16_t data[kDataLength] = {0}; + AudioFeatures features; + double sp[kMaxNumFrames]; + while (fread(data, sizeof(int16_t), kDataLength, pcm_file) == kDataLength) { + audioproc.ExtractFeatures(data, kDataLength, &features); + if (features.num_frames > 0) { + ASSERT_LT(features.num_frames, kMaxNumFrames); + // Read reference values. + const size_t num_frames = features.num_frames; + ASSERT_EQ(num_frames, fread(sp, sizeof(sp[0]), num_frames, peak_file)); + for (size_t n = 0; n < features.num_frames; n++) + EXPECT_NEAR(features.spectral_peak[n], sp[n], 3); + } + } + + fclose(peak_file); + fclose(pcm_file); +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/vad_circular_buffer.cc b/webrtc/modules/audio_processing/vad/vad_circular_buffer.cc new file mode 100644 index 0000000000..d337893c45 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/vad_circular_buffer.cc @@ -0,0 +1,138 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/vad_circular_buffer.h" + +#include <assert.h> +#include <stdlib.h> + +namespace webrtc { + +VadCircularBuffer::VadCircularBuffer(int buffer_size) + : buffer_(new double[buffer_size]), + is_full_(false), + index_(0), + buffer_size_(buffer_size), + sum_(0) { +} + +VadCircularBuffer::~VadCircularBuffer() { +} + +void VadCircularBuffer::Reset() { + is_full_ = false; + index_ = 0; + sum_ = 0; +} + +VadCircularBuffer* VadCircularBuffer::Create(int buffer_size) { + if (buffer_size <= 0) + return NULL; + return new VadCircularBuffer(buffer_size); +} + +double VadCircularBuffer::Oldest() const { + if (!is_full_) + return buffer_[0]; + else + return buffer_[index_]; +} + +double VadCircularBuffer::Mean() { + double m; + if (is_full_) { + m = sum_ / buffer_size_; + } else { + if (index_ > 0) + m = sum_ / index_; + else + m = 0; + } + return m; +} + +void VadCircularBuffer::Insert(double value) { + if (is_full_) { + sum_ -= buffer_[index_]; + } + sum_ += value; + buffer_[index_] = value; + index_++; + if (index_ >= buffer_size_) { + is_full_ = true; + index_ = 0; + } +} +int VadCircularBuffer::BufferLevel() { + if (is_full_) + return buffer_size_; + return index_; +} + +int VadCircularBuffer::Get(int index, double* value) const { + int err = ConvertToLinearIndex(&index); + if (err < 0) + return -1; + *value = buffer_[index]; + return 0; +} + +int VadCircularBuffer::Set(int index, double value) { + int err = ConvertToLinearIndex(&index); + if (err < 0) + return -1; + + sum_ -= buffer_[index]; + buffer_[index] = value; + sum_ += value; + return 0; +} + +int VadCircularBuffer::ConvertToLinearIndex(int* index) const { + if (*index < 0 || *index >= buffer_size_) + return -1; + + if (!is_full_ && *index >= index_) + return -1; + + *index = index_ - 1 - *index; + if (*index < 0) + *index += buffer_size_; + return 0; +} + +int VadCircularBuffer::RemoveTransient(int width_threshold, + double val_threshold) { + if (!is_full_ && index_ < width_threshold + 2) + return 0; + + int index_1 = 0; + int index_2 = width_threshold + 1; + double v = 0; + if (Get(index_1, &v) < 0) + return -1; + if (v < val_threshold) { + Set(index_1, 0); + int index; + for (index = index_2; index > index_1; index--) { + if (Get(index, &v) < 0) + return -1; + if (v < val_threshold) + break; + } + for (; index > index_1; index--) { + if (Set(index, 0.0) < 0) + return -1; + } + } + return 0; +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/vad_circular_buffer.h b/webrtc/modules/audio_processing/vad/vad_circular_buffer.h new file mode 100644 index 0000000000..5238f77257 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/vad_circular_buffer.h @@ -0,0 +1,69 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VAD_CIRCULAR_BUFFER_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VAD_CIRCULAR_BUFFER_H_ + +#include "webrtc/base/scoped_ptr.h" + +namespace webrtc { + +// A circular buffer tailored to the need of this project. It stores last +// K samples of the input, and keeps track of the mean of the last samples. +// +// It is used in class "PitchBasedActivity" to keep track of posterior +// probabilities in the past few seconds. The posterior probabilities are used +// to recursively update prior probabilities. +class VadCircularBuffer { + public: + static VadCircularBuffer* Create(int buffer_size); + ~VadCircularBuffer(); + + // If buffer is wrapped around. + bool is_full() const { return is_full_; } + // Get the oldest entry in the buffer. + double Oldest() const; + // Insert new value into the buffer. + void Insert(double value); + // Reset buffer, forget the past, start fresh. + void Reset(); + + // The mean value of the elements in the buffer. The return value is zero if + // buffer is empty, i.e. no value is inserted. + double Mean(); + // Remove transients. If the values exceed |val_threshold| for a period + // shorter then or equal to |width_threshold|, then that period is considered + // transient and set to zero. + int RemoveTransient(int width_threshold, double val_threshold); + + private: + explicit VadCircularBuffer(int buffer_size); + // Get previous values. |index = 0| corresponds to the most recent + // insertion. |index = 1| is the one before the most recent insertion, and + // so on. + int Get(int index, double* value) const; + // Set a given position to |value|. |index| is interpreted as above. + int Set(int index, double value); + // Return the number of valid elements in the buffer. + int BufferLevel(); + + // Convert an index with the interpretation as get() method to the + // corresponding linear index. + int ConvertToLinearIndex(int* index) const; + + rtc::scoped_ptr<double[]> buffer_; + bool is_full_; + int index_; + int buffer_size_; + double sum_; +}; + +} // namespace webrtc +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VAD_CIRCULAR_BUFFER_H_ diff --git a/webrtc/modules/audio_processing/vad/vad_circular_buffer_unittest.cc b/webrtc/modules/audio_processing/vad/vad_circular_buffer_unittest.cc new file mode 100644 index 0000000000..11945e042c --- /dev/null +++ b/webrtc/modules/audio_processing/vad/vad_circular_buffer_unittest.cc @@ -0,0 +1,133 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/vad_circular_buffer.h" + +#include <stdio.h> + +#include "testing/gtest/include/gtest/gtest.h" +#include "webrtc/base/scoped_ptr.h" + +namespace webrtc { + +static const int kWidthThreshold = 7; +static const double kValThreshold = 1.0; +static const int kLongBuffSize = 100; +static const int kShortBuffSize = 10; + +static void InsertSequentially(int k, VadCircularBuffer* circular_buffer) { + double mean_val; + for (int n = 1; n <= k; n++) { + EXPECT_TRUE(!circular_buffer->is_full()); + circular_buffer->Insert(n); + mean_val = circular_buffer->Mean(); + EXPECT_EQ((n + 1.0) / 2., mean_val); + } +} + +static void Insert(double value, + int num_insertion, + VadCircularBuffer* circular_buffer) { + for (int n = 0; n < num_insertion; n++) + circular_buffer->Insert(value); +} + +static void InsertZeros(int num_zeros, VadCircularBuffer* circular_buffer) { + Insert(0.0, num_zeros, circular_buffer); +} + +TEST(VadCircularBufferTest, GeneralTest) { + rtc::scoped_ptr<VadCircularBuffer> circular_buffer( + VadCircularBuffer::Create(kShortBuffSize)); + double mean_val; + + // Mean should return zero if nothing is inserted. + mean_val = circular_buffer->Mean(); + EXPECT_DOUBLE_EQ(0.0, mean_val); + InsertSequentially(kShortBuffSize, circular_buffer.get()); + + // Should be full. + EXPECT_TRUE(circular_buffer->is_full()); + // Correct update after being full. + for (int n = 1; n < kShortBuffSize; n++) { + circular_buffer->Insert(n); + mean_val = circular_buffer->Mean(); + EXPECT_DOUBLE_EQ((kShortBuffSize + 1.) / 2., mean_val); + EXPECT_TRUE(circular_buffer->is_full()); + } + + // Check reset. This should be like starting fresh. + circular_buffer->Reset(); + mean_val = circular_buffer->Mean(); + EXPECT_DOUBLE_EQ(0, mean_val); + InsertSequentially(kShortBuffSize, circular_buffer.get()); + EXPECT_TRUE(circular_buffer->is_full()); +} + +TEST(VadCircularBufferTest, TransientsRemoval) { + rtc::scoped_ptr<VadCircularBuffer> circular_buffer( + VadCircularBuffer::Create(kLongBuffSize)); + // Let the first transient be in wrap-around. + InsertZeros(kLongBuffSize - kWidthThreshold / 2, circular_buffer.get()); + + double push_val = kValThreshold; + double mean_val; + for (int k = kWidthThreshold; k >= 1; k--) { + Insert(push_val, k, circular_buffer.get()); + circular_buffer->Insert(0); + mean_val = circular_buffer->Mean(); + EXPECT_DOUBLE_EQ(k * push_val / kLongBuffSize, mean_val); + circular_buffer->RemoveTransient(kWidthThreshold, kValThreshold); + mean_val = circular_buffer->Mean(); + EXPECT_DOUBLE_EQ(0, mean_val); + } +} + +TEST(VadCircularBufferTest, TransientDetection) { + rtc::scoped_ptr<VadCircularBuffer> circular_buffer( + VadCircularBuffer::Create(kLongBuffSize)); + // Let the first transient be in wrap-around. + int num_insertion = kLongBuffSize - kWidthThreshold / 2; + InsertZeros(num_insertion, circular_buffer.get()); + + double push_val = 2; + // This is longer than a transient and shouldn't be removed. + int num_non_zero_elements = kWidthThreshold + 1; + Insert(push_val, num_non_zero_elements, circular_buffer.get()); + + double mean_val = circular_buffer->Mean(); + EXPECT_DOUBLE_EQ(num_non_zero_elements * push_val / kLongBuffSize, mean_val); + circular_buffer->Insert(0); + EXPECT_EQ(0, + circular_buffer->RemoveTransient(kWidthThreshold, kValThreshold)); + mean_val = circular_buffer->Mean(); + EXPECT_DOUBLE_EQ(num_non_zero_elements * push_val / kLongBuffSize, mean_val); + + // A transient right after a non-transient, should be removed and mean is + // not changed. + num_insertion = 3; + Insert(push_val, num_insertion, circular_buffer.get()); + circular_buffer->Insert(0); + EXPECT_EQ(0, + circular_buffer->RemoveTransient(kWidthThreshold, kValThreshold)); + mean_val = circular_buffer->Mean(); + EXPECT_DOUBLE_EQ(num_non_zero_elements * push_val / kLongBuffSize, mean_val); + + // Last input is larger than threshold, although the sequence is short but + // it shouldn't be considered transient. + Insert(push_val, num_insertion, circular_buffer.get()); + num_non_zero_elements += num_insertion; + EXPECT_EQ(0, + circular_buffer->RemoveTransient(kWidthThreshold, kValThreshold)); + mean_val = circular_buffer->Mean(); + EXPECT_DOUBLE_EQ(num_non_zero_elements * push_val / kLongBuffSize, mean_val); +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/voice_activity_detector.cc b/webrtc/modules/audio_processing/vad/voice_activity_detector.cc new file mode 100644 index 0000000000..ef56a3574c --- /dev/null +++ b/webrtc/modules/audio_processing/vad/voice_activity_detector.cc @@ -0,0 +1,85 @@ +/* + * Copyright (c) 2015 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/voice_activity_detector.h" + +#include <algorithm> + +#include "webrtc/base/checks.h" + +namespace webrtc { +namespace { + +const size_t kMaxLength = 320; +const int kNumChannels = 1; + +const double kDefaultVoiceValue = 1.0; +const double kNeutralProbability = 0.5; +const double kLowProbability = 0.01; + +} // namespace + +VoiceActivityDetector::VoiceActivityDetector() + : last_voice_probability_(kDefaultVoiceValue), + standalone_vad_(StandaloneVad::Create()) { +} + +// Because ISAC has a different chunk length, it updates +// |chunkwise_voice_probabilities_| and |chunkwise_rms_| when there is new data. +// Otherwise it clears them. +void VoiceActivityDetector::ProcessChunk(const int16_t* audio, + size_t length, + int sample_rate_hz) { + RTC_DCHECK_EQ(static_cast<int>(length), sample_rate_hz / 100); + RTC_DCHECK_LE(length, kMaxLength); + // Resample to the required rate. + const int16_t* resampled_ptr = audio; + if (sample_rate_hz != kSampleRateHz) { + RTC_CHECK_EQ( + resampler_.ResetIfNeeded(sample_rate_hz, kSampleRateHz, kNumChannels), + 0); + resampler_.Push(audio, length, resampled_, kLength10Ms, length); + resampled_ptr = resampled_; + } + RTC_DCHECK_EQ(length, kLength10Ms); + + // Each chunk needs to be passed into |standalone_vad_|, because internally it + // buffers the audio and processes it all at once when GetActivity() is + // called. + RTC_CHECK_EQ(standalone_vad_->AddAudio(resampled_ptr, length), 0); + + audio_processing_.ExtractFeatures(resampled_ptr, length, &features_); + + chunkwise_voice_probabilities_.resize(features_.num_frames); + chunkwise_rms_.resize(features_.num_frames); + std::copy(features_.rms, features_.rms + chunkwise_rms_.size(), + chunkwise_rms_.begin()); + if (features_.num_frames > 0) { + if (features_.silence) { + // The other features are invalid, so set the voice probabilities to an + // arbitrary low value. + std::fill(chunkwise_voice_probabilities_.begin(), + chunkwise_voice_probabilities_.end(), kLowProbability); + } else { + std::fill(chunkwise_voice_probabilities_.begin(), + chunkwise_voice_probabilities_.end(), kNeutralProbability); + RTC_CHECK_GE( + standalone_vad_->GetActivity(&chunkwise_voice_probabilities_[0], + chunkwise_voice_probabilities_.size()), + 0); + RTC_CHECK_GE(pitch_based_vad_.VoicingProbability( + features_, &chunkwise_voice_probabilities_[0]), + 0); + } + last_voice_probability_ = chunkwise_voice_probabilities_.back(); + } +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/voice_activity_detector.h b/webrtc/modules/audio_processing/vad/voice_activity_detector.h new file mode 100644 index 0000000000..e2dcf022a9 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/voice_activity_detector.h @@ -0,0 +1,70 @@ +/* + * Copyright (c) 2015 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VOICE_ACTIVITY_DETECTOR_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VOICE_ACTIVITY_DETECTOR_H_ + +#include <vector> + +#include "webrtc/base/scoped_ptr.h" +#include "webrtc/common_audio/resampler/include/resampler.h" +#include "webrtc/modules/audio_processing/vad/vad_audio_proc.h" +#include "webrtc/modules/audio_processing/vad/common.h" +#include "webrtc/modules/audio_processing/vad/pitch_based_vad.h" +#include "webrtc/modules/audio_processing/vad/standalone_vad.h" + +namespace webrtc { + +// A Voice Activity Detector (VAD) that combines the voice probability from the +// StandaloneVad and PitchBasedVad to get a more robust estimation. +class VoiceActivityDetector { + public: + VoiceActivityDetector(); + + // Processes each audio chunk and estimates the voice probability. The maximum + // supported sample rate is 32kHz. + // TODO(aluebs): Change |length| to size_t. + void ProcessChunk(const int16_t* audio, size_t length, int sample_rate_hz); + + // Returns a vector of voice probabilities for each chunk. It can be empty for + // some chunks, but it catches up afterwards returning multiple values at + // once. + const std::vector<double>& chunkwise_voice_probabilities() const { + return chunkwise_voice_probabilities_; + } + + // Returns a vector of RMS values for each chunk. It has the same length as + // chunkwise_voice_probabilities(). + const std::vector<double>& chunkwise_rms() const { return chunkwise_rms_; } + + // Returns the last voice probability, regardless of the internal + // implementation, although it has a few chunks of delay. + float last_voice_probability() const { return last_voice_probability_; } + + private: + // TODO(aluebs): Change these to float. + std::vector<double> chunkwise_voice_probabilities_; + std::vector<double> chunkwise_rms_; + + float last_voice_probability_; + + Resampler resampler_; + VadAudioProc audio_processing_; + + rtc::scoped_ptr<StandaloneVad> standalone_vad_; + PitchBasedVad pitch_based_vad_; + + int16_t resampled_[kLength10Ms]; + AudioFeatures features_; +}; + +} // namespace webrtc + +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VOICE_ACTIVITY_DETECTOR_H_ diff --git a/webrtc/modules/audio_processing/vad/voice_activity_detector_unittest.cc b/webrtc/modules/audio_processing/vad/voice_activity_detector_unittest.cc new file mode 100644 index 0000000000..f4ee17760e --- /dev/null +++ b/webrtc/modules/audio_processing/vad/voice_activity_detector_unittest.cc @@ -0,0 +1,168 @@ +/* + * Copyright (c) 2015 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/vad/voice_activity_detector.h" + +#include <algorithm> +#include <vector> + +#include "testing/gtest/include/gtest/gtest.h" +#include "webrtc/test/testsupport/fileutils.h" + +namespace webrtc { +namespace { + +const int kStartTimeSec = 16; +const float kMeanSpeechProbability = 0.3f; +const float kMaxNoiseProbability = 0.1f; +const size_t kNumChunks = 300u; +const size_t kNumChunksPerIsacBlock = 3; + +void GenerateNoise(std::vector<int16_t>* data) { + for (size_t i = 0; i < data->size(); ++i) { + // std::rand returns between 0 and RAND_MAX, but this will work because it + // wraps into some random place. + (*data)[i] = std::rand(); + } +} + +} // namespace + +TEST(VoiceActivityDetectorTest, ConstructorSetsDefaultValues) { + const float kDefaultVoiceValue = 1.f; + + VoiceActivityDetector vad; + + std::vector<double> p = vad.chunkwise_voice_probabilities(); + std::vector<double> rms = vad.chunkwise_rms(); + + EXPECT_EQ(p.size(), 0u); + EXPECT_EQ(rms.size(), 0u); + + EXPECT_FLOAT_EQ(vad.last_voice_probability(), kDefaultVoiceValue); +} + +TEST(VoiceActivityDetectorTest, Speech16kHzHasHighVoiceProbabilities) { + const int kSampleRateHz = 16000; + const int kLength10Ms = kSampleRateHz / 100; + + VoiceActivityDetector vad; + + std::vector<int16_t> data(kLength10Ms); + float mean_probability = 0.f; + + FILE* pcm_file = + fopen(test::ResourcePath("audio_processing/transient/audio16kHz", "pcm") + .c_str(), + "rb"); + ASSERT_TRUE(pcm_file != nullptr); + // The silences in the file are skipped to get a more robust voice probability + // for speech. + ASSERT_EQ(fseek(pcm_file, kStartTimeSec * kSampleRateHz * sizeof(data[0]), + SEEK_SET), + 0); + + size_t num_chunks = 0; + while (fread(&data[0], sizeof(data[0]), data.size(), pcm_file) == + data.size()) { + vad.ProcessChunk(&data[0], data.size(), kSampleRateHz); + + mean_probability += vad.last_voice_probability(); + + ++num_chunks; + } + + mean_probability /= num_chunks; + + EXPECT_GT(mean_probability, kMeanSpeechProbability); +} + +TEST(VoiceActivityDetectorTest, Speech32kHzHasHighVoiceProbabilities) { + const int kSampleRateHz = 32000; + const int kLength10Ms = kSampleRateHz / 100; + + VoiceActivityDetector vad; + + std::vector<int16_t> data(kLength10Ms); + float mean_probability = 0.f; + + FILE* pcm_file = + fopen(test::ResourcePath("audio_processing/transient/audio32kHz", "pcm") + .c_str(), + "rb"); + ASSERT_TRUE(pcm_file != nullptr); + // The silences in the file are skipped to get a more robust voice probability + // for speech. + ASSERT_EQ(fseek(pcm_file, kStartTimeSec * kSampleRateHz * sizeof(data[0]), + SEEK_SET), + 0); + + size_t num_chunks = 0; + while (fread(&data[0], sizeof(data[0]), data.size(), pcm_file) == + data.size()) { + vad.ProcessChunk(&data[0], data.size(), kSampleRateHz); + + mean_probability += vad.last_voice_probability(); + + ++num_chunks; + } + + mean_probability /= num_chunks; + + EXPECT_GT(mean_probability, kMeanSpeechProbability); +} + +TEST(VoiceActivityDetectorTest, Noise16kHzHasLowVoiceProbabilities) { + VoiceActivityDetector vad; + + std::vector<int16_t> data(kLength10Ms); + float max_probability = 0.f; + + std::srand(42); + + for (size_t i = 0; i < kNumChunks; ++i) { + GenerateNoise(&data); + + vad.ProcessChunk(&data[0], data.size(), kSampleRateHz); + + // Before the |vad has enough data to process an ISAC block it will return + // the default value, 1.f, which would ruin the |max_probability| value. + if (i > kNumChunksPerIsacBlock) { + max_probability = std::max(max_probability, vad.last_voice_probability()); + } + } + + EXPECT_LT(max_probability, kMaxNoiseProbability); +} + +TEST(VoiceActivityDetectorTest, Noise32kHzHasLowVoiceProbabilities) { + VoiceActivityDetector vad; + + std::vector<int16_t> data(2 * kLength10Ms); + float max_probability = 0.f; + + std::srand(42); + + for (size_t i = 0; i < kNumChunks; ++i) { + GenerateNoise(&data); + + vad.ProcessChunk(&data[0], data.size(), 2 * kSampleRateHz); + + // Before the |vad has enough data to process an ISAC block it will return + // the default value, 1.f, which would ruin the |max_probability| value. + if (i > kNumChunksPerIsacBlock) { + max_probability = std::max(max_probability, vad.last_voice_probability()); + } + } + + EXPECT_LT(max_probability, kMaxNoiseProbability); +} + +} // namespace webrtc diff --git a/webrtc/modules/audio_processing/vad/voice_gmm_tables.h b/webrtc/modules/audio_processing/vad/voice_gmm_tables.h new file mode 100644 index 0000000000..2f247c3798 --- /dev/null +++ b/webrtc/modules/audio_processing/vad/voice_gmm_tables.h @@ -0,0 +1,85 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +// GMM tables for active segments. Generated by MakeGmmTables.m. + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VOICE_GMM_TABLES_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VOICE_GMM_TABLES_H_ + +static const int kVoiceGmmNumMixtures = 12; +static const int kVoiceGmmDim = 3; + +static const double + kVoiceGmmCovarInverse[kVoiceGmmNumMixtures][kVoiceGmmDim][kVoiceGmmDim] = { + {{1.83673825579513e+00, -8.09791637570095e-04, 4.60106414365986e-03}, + {-8.09791637570095e-04, 8.89351738394608e-04, -9.80188953277734e-04}, + {4.60106414365986e-03, -9.80188953277734e-04, 1.38706060206582e-03}}, + {{6.76228912850703e+01, -1.98893120119660e-02, -3.53548357253551e-03}, + {-1.98893120119660e-02, 3.96216858500530e-05, -4.08492938394097e-05}, + {-3.53548357253551e-03, -4.08492938394097e-05, 9.31864352856416e-04}}, + {{9.98612435944558e+00, -5.27880954316893e-03, -6.30342541619017e-03}, + {-5.27880954316893e-03, 4.54359480225226e-05, 6.30804591626044e-05}, + {-6.30342541619017e-03, 6.30804591626044e-05, 5.36466441382942e-04}}, + {{3.39917474216349e+01, -1.56213579433191e-03, -4.01459014990225e-02}, + {-1.56213579433191e-03, 6.40415424897724e-05, 6.20076342427833e-05}, + {-4.01459014990225e-02, 6.20076342427833e-05, 3.51199070103063e-03}}, + {{1.34545062271428e+01, -7.94513610147144e-03, -5.34401019341728e-02}, + {-7.94513610147144e-03, 1.16511820098649e-04, 4.66063702069293e-05}, + {-5.34401019341728e-02, 4.66063702069293e-05, 2.72354323774163e-03}}, + {{1.08557844314806e+02, -1.54885805673668e-02, -1.88029692674851e-02}, + {-1.54885805673668e-02, 1.16404042786406e-04, 6.45579292702802e-06}, + {-1.88029692674851e-02, 6.45579292702802e-06, 4.32330478391416e-04}}, + {{8.22940066541450e+01, -1.15903110231303e-02, -4.92166764865343e-02}, + {-1.15903110231303e-02, 7.42510742165261e-05, 3.73007314191290e-06}, + {-4.92166764865343e-02, 3.73007314191290e-06, 3.64005221593244e-03}}, + {{2.31133605685660e+00, -7.83261568950254e-04, 7.45744012346313e-04}, + {-7.83261568950254e-04, 1.29460648214142e-05, -2.22774455093730e-06}, + {7.45744012346313e-04, -2.22774455093730e-06, 1.05117294093010e-04}}, + {{3.78767849189611e+02, 1.57759761011568e-03, -2.08551217988774e-02}, + {1.57759761011568e-03, 4.76066236886865e-05, -2.33977412299324e-05}, + {-2.08551217988774e-02, -2.33977412299324e-05, 5.24261005371196e-04}}, + {{6.98580096506135e-01, -5.13850255217378e-04, -4.01124551717056e-04}, + {-5.13850255217378e-04, 1.40501021984840e-06, -2.09496928716569e-06}, + {-4.01124551717056e-04, -2.09496928716569e-06, 2.82879357740037e-04}}, + {{2.62770945162399e+00, -2.31825753241430e-03, -5.30447217466318e-03}, + {-2.31825753241430e-03, 4.59108572227649e-05, 7.67631886355405e-05}, + {-5.30447217466318e-03, 7.67631886355405e-05, 2.28521601674098e-03}}, + {{1.89940391362152e+02, -4.23280856852379e-03, -2.70608873541399e-02}, + {-4.23280856852379e-03, 6.77547582742563e-05, 2.69154203800467e-05}, + {-2.70608873541399e-02, 2.69154203800467e-05, 3.88574543373470e-03}}}; + +static const double kVoiceGmmMean[kVoiceGmmNumMixtures][kVoiceGmmDim] = { + {-2.15020241646536e+00, 4.97079062999877e+02, 4.77078119504505e+02}, + {-8.92097680029190e-01, 5.92064964199921e+02, 1.81045145941059e+02}, + {-1.29435784144398e+00, 4.98450293410611e+02, 1.71991263804064e+02}, + {-1.03925228397884e+00, 4.99511274321571e+02, 1.05838336539105e+02}, + {-1.29229047206129e+00, 4.15026762566707e+02, 1.12861119017125e+02}, + {-7.88748114599810e-01, 4.48739336688113e+02, 1.89784216956337e+02}, + {-8.77777402332642e-01, 4.86620285054533e+02, 1.13477708016491e+02}, + {-2.06465957063057e+00, 6.33385049870607e+02, 2.32758546796149e+02}, + {-6.98893789231685e-01, 5.93622051503385e+02, 1.92536982473203e+02}, + {-2.55901217508894e+00, 1.55914919756205e+03, 1.39769980835570e+02}, + {-1.92070024165837e+00, 4.87983940444185e+02, 1.02745468128289e+02}, + {-7.29187507662854e-01, 5.22717685022855e+02, 1.16377942283991e+02}}; + +static const double kVoiceGmmWeights[kVoiceGmmNumMixtures] = { + -1.39789694361035e+01, + -1.19527720202104e+01, + -1.32396317929055e+01, + -1.09436815209238e+01, + -1.13440027478149e+01, + -1.12200721834504e+01, + -1.02537324043693e+01, + -1.60789861938302e+01, + -1.03394494048344e+01, + -1.83207938586818e+01, + -1.31186044948288e+01, + -9.52479998673554e+00}; +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_VAD_VOICE_GMM_TABLES_H_ |