1 /*
2 * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include "webrtc/modules/audio_processing/vad/vad_audio_proc.h"
12
13 #include <math.h>
14 #include <stdio.h>
15
16 #include "webrtc/common_audio/fft4g.h"
17 #include "webrtc/modules/audio_processing/vad/vad_audio_proc_internal.h"
18 #include "webrtc/modules/audio_processing/vad/pitch_internal.h"
19 #include "webrtc/modules/audio_processing/vad/pole_zero_filter.h"
20 extern "C" {
21 #include "webrtc/modules/audio_coding/codecs/isac/main/source/codec.h"
22 #include "webrtc/modules/audio_coding/codecs/isac/main/source/lpc_analysis.h"
23 #include "webrtc/modules/audio_coding/codecs/isac/main/source/pitch_estimator.h"
24 #include "webrtc/modules/audio_coding/codecs/isac/main/source/structs.h"
25 }
26 #include "webrtc/modules/include/module_common_types.h"
27
28 namespace webrtc {
29
30 // The following structures are declared anonymous in iSAC's structs.h. To
31 // forward declare them, we use this derived class trick.
32 struct VadAudioProc::PitchAnalysisStruct : public ::PitchAnalysisStruct {};
33 struct VadAudioProc::PreFiltBankstr : public ::PreFiltBankstr {};
34
35 static const float kFrequencyResolution =
36 kSampleRateHz / static_cast<float>(VadAudioProc::kDftSize);
37 static const int kSilenceRms = 5;
38
39 // TODO(turajs): Make a Create or Init for VadAudioProc.
VadAudioProc()40 VadAudioProc::VadAudioProc()
41 : audio_buffer_(),
42 num_buffer_samples_(kNumPastSignalSamples),
43 log_old_gain_(-2),
44 old_lag_(50), // Arbitrary but valid as pitch-lag (in samples).
45 pitch_analysis_handle_(new PitchAnalysisStruct),
46 pre_filter_handle_(new PreFiltBankstr),
47 high_pass_filter_(PoleZeroFilter::Create(kCoeffNumerator,
48 kFilterOrder,
49 kCoeffDenominator,
50 kFilterOrder)) {
51 static_assert(kNumPastSignalSamples + kNumSubframeSamples ==
52 sizeof(kLpcAnalWin) / sizeof(kLpcAnalWin[0]),
53 "lpc analysis window incorrect size");
54 static_assert(kLpcOrder + 1 == sizeof(kCorrWeight) / sizeof(kCorrWeight[0]),
55 "correlation weight incorrect size");
56
57 // TODO(turajs): Are we doing too much in the constructor?
58 float data[kDftSize];
59 // Make FFT to initialize.
60 ip_[0] = 0;
61 WebRtc_rdft(kDftSize, 1, data, ip_, w_fft_);
62 // TODO(turajs): Need to initialize high-pass filter.
63
64 // Initialize iSAC components.
65 WebRtcIsac_InitPreFilterbank(pre_filter_handle_.get());
66 WebRtcIsac_InitPitchAnalysis(pitch_analysis_handle_.get());
67 }
68
~VadAudioProc()69 VadAudioProc::~VadAudioProc() {
70 }
71
ResetBuffer()72 void VadAudioProc::ResetBuffer() {
73 memcpy(audio_buffer_, &audio_buffer_[kNumSamplesToProcess],
74 sizeof(audio_buffer_[0]) * kNumPastSignalSamples);
75 num_buffer_samples_ = kNumPastSignalSamples;
76 }
77
ExtractFeatures(const int16_t * frame,size_t length,AudioFeatures * features)78 int VadAudioProc::ExtractFeatures(const int16_t* frame,
79 size_t length,
80 AudioFeatures* features) {
81 features->num_frames = 0;
82 if (length != kNumSubframeSamples) {
83 return -1;
84 }
85
86 // High-pass filter to remove the DC component and very low frequency content.
87 // We have experienced that this high-pass filtering improves voice/non-voiced
88 // classification.
89 if (high_pass_filter_->Filter(frame, kNumSubframeSamples,
90 &audio_buffer_[num_buffer_samples_]) != 0) {
91 return -1;
92 }
93
94 num_buffer_samples_ += kNumSubframeSamples;
95 if (num_buffer_samples_ < kBufferLength) {
96 return 0;
97 }
98 assert(num_buffer_samples_ == kBufferLength);
99 features->num_frames = kNum10msSubframes;
100 features->silence = false;
101
102 Rms(features->rms, kMaxNumFrames);
103 for (size_t i = 0; i < kNum10msSubframes; ++i) {
104 if (features->rms[i] < kSilenceRms) {
105 // PitchAnalysis can cause NaNs in the pitch gain if it's fed silence.
106 // Bail out here instead.
107 features->silence = true;
108 ResetBuffer();
109 return 0;
110 }
111 }
112
113 PitchAnalysis(features->log_pitch_gain, features->pitch_lag_hz,
114 kMaxNumFrames);
115 FindFirstSpectralPeaks(features->spectral_peak, kMaxNumFrames);
116 ResetBuffer();
117 return 0;
118 }
119
120 // Computes |kLpcOrder + 1| correlation coefficients.
SubframeCorrelation(double * corr,size_t length_corr,size_t subframe_index)121 void VadAudioProc::SubframeCorrelation(double* corr,
122 size_t length_corr,
123 size_t subframe_index) {
124 assert(length_corr >= kLpcOrder + 1);
125 double windowed_audio[kNumSubframeSamples + kNumPastSignalSamples];
126 size_t buffer_index = subframe_index * kNumSubframeSamples;
127
128 for (size_t n = 0; n < kNumSubframeSamples + kNumPastSignalSamples; n++)
129 windowed_audio[n] = audio_buffer_[buffer_index++] * kLpcAnalWin[n];
130
131 WebRtcIsac_AutoCorr(corr, windowed_audio,
132 kNumSubframeSamples + kNumPastSignalSamples, kLpcOrder);
133 }
134
135 // Compute |kNum10msSubframes| sets of LPC coefficients, one per 10 ms input.
136 // The analysis window is 15 ms long and it is centered on the first half of
137 // each 10ms sub-frame. This is equivalent to computing LPC coefficients for the
138 // first half of each 10 ms subframe.
GetLpcPolynomials(double * lpc,size_t length_lpc)139 void VadAudioProc::GetLpcPolynomials(double* lpc, size_t length_lpc) {
140 assert(length_lpc >= kNum10msSubframes * (kLpcOrder + 1));
141 double corr[kLpcOrder + 1];
142 double reflec_coeff[kLpcOrder];
143 for (size_t i = 0, offset_lpc = 0; i < kNum10msSubframes;
144 i++, offset_lpc += kLpcOrder + 1) {
145 SubframeCorrelation(corr, kLpcOrder + 1, i);
146 corr[0] *= 1.0001;
147 // This makes Lev-Durb a bit more stable.
148 for (size_t k = 0; k < kLpcOrder + 1; k++) {
149 corr[k] *= kCorrWeight[k];
150 }
151 WebRtcIsac_LevDurb(&lpc[offset_lpc], reflec_coeff, corr, kLpcOrder);
152 }
153 }
154
155 // Fit a second order curve to these 3 points and find the location of the
156 // extremum. The points are inverted before curve fitting.
QuadraticInterpolation(float prev_val,float curr_val,float next_val)157 static float QuadraticInterpolation(float prev_val,
158 float curr_val,
159 float next_val) {
160 // Doing the interpolation in |1 / A(z)|^2.
161 float fractional_index = 0;
162 next_val = 1.0f / next_val;
163 prev_val = 1.0f / prev_val;
164 curr_val = 1.0f / curr_val;
165
166 fractional_index =
167 -(next_val - prev_val) * 0.5f / (next_val + prev_val - 2.f * curr_val);
168 assert(fabs(fractional_index) < 1);
169 return fractional_index;
170 }
171
172 // 1 / A(z), where A(z) is defined by |lpc| is a model of the spectral envelope
173 // of the input signal. The local maximum of the spectral envelope corresponds
174 // with the local minimum of A(z). It saves complexity, as we save one
175 // inversion. Furthermore, we find the first local maximum of magnitude squared,
176 // to save on one square root.
FindFirstSpectralPeaks(double * f_peak,size_t length_f_peak)177 void VadAudioProc::FindFirstSpectralPeaks(double* f_peak,
178 size_t length_f_peak) {
179 assert(length_f_peak >= kNum10msSubframes);
180 double lpc[kNum10msSubframes * (kLpcOrder + 1)];
181 // For all sub-frames.
182 GetLpcPolynomials(lpc, kNum10msSubframes * (kLpcOrder + 1));
183
184 const size_t kNumDftCoefficients = kDftSize / 2 + 1;
185 float data[kDftSize];
186
187 for (size_t i = 0; i < kNum10msSubframes; i++) {
188 // Convert to float with zero pad.
189 memset(data, 0, sizeof(data));
190 for (size_t n = 0; n < kLpcOrder + 1; n++) {
191 data[n] = static_cast<float>(lpc[i * (kLpcOrder + 1) + n]);
192 }
193 // Transform to frequency domain.
194 WebRtc_rdft(kDftSize, 1, data, ip_, w_fft_);
195
196 size_t index_peak = 0;
197 float prev_magn_sqr = data[0] * data[0];
198 float curr_magn_sqr = data[2] * data[2] + data[3] * data[3];
199 float next_magn_sqr;
200 bool found_peak = false;
201 for (size_t n = 2; n < kNumDftCoefficients - 1; n++) {
202 next_magn_sqr =
203 data[2 * n] * data[2 * n] + data[2 * n + 1] * data[2 * n + 1];
204 if (curr_magn_sqr < prev_magn_sqr && curr_magn_sqr < next_magn_sqr) {
205 found_peak = true;
206 index_peak = n - 1;
207 break;
208 }
209 prev_magn_sqr = curr_magn_sqr;
210 curr_magn_sqr = next_magn_sqr;
211 }
212 float fractional_index = 0;
213 if (!found_peak) {
214 // Checking if |kNumDftCoefficients - 1| is the local minimum.
215 next_magn_sqr = data[1] * data[1];
216 if (curr_magn_sqr < prev_magn_sqr && curr_magn_sqr < next_magn_sqr) {
217 index_peak = kNumDftCoefficients - 1;
218 }
219 } else {
220 // A peak is found, do a simple quadratic interpolation to get a more
221 // accurate estimate of the peak location.
222 fractional_index =
223 QuadraticInterpolation(prev_magn_sqr, curr_magn_sqr, next_magn_sqr);
224 }
225 f_peak[i] = (index_peak + fractional_index) * kFrequencyResolution;
226 }
227 }
228
229 // Using iSAC functions to estimate pitch gains & lags.
PitchAnalysis(double * log_pitch_gains,double * pitch_lags_hz,size_t length)230 void VadAudioProc::PitchAnalysis(double* log_pitch_gains,
231 double* pitch_lags_hz,
232 size_t length) {
233 // TODO(turajs): This can be "imported" from iSAC & and the next two
234 // constants.
235 assert(length >= kNum10msSubframes);
236 const int kNumPitchSubframes = 4;
237 double gains[kNumPitchSubframes];
238 double lags[kNumPitchSubframes];
239
240 const int kNumSubbandFrameSamples = 240;
241 const int kNumLookaheadSamples = 24;
242
243 float lower[kNumSubbandFrameSamples];
244 float upper[kNumSubbandFrameSamples];
245 double lower_lookahead[kNumSubbandFrameSamples];
246 double upper_lookahead[kNumSubbandFrameSamples];
247 double lower_lookahead_pre_filter[kNumSubbandFrameSamples +
248 kNumLookaheadSamples];
249
250 // Split signal to lower and upper bands
251 WebRtcIsac_SplitAndFilterFloat(&audio_buffer_[kNumPastSignalSamples], lower,
252 upper, lower_lookahead, upper_lookahead,
253 pre_filter_handle_.get());
254 WebRtcIsac_PitchAnalysis(lower_lookahead, lower_lookahead_pre_filter,
255 pitch_analysis_handle_.get(), lags, gains);
256
257 // Lags are computed on lower-band signal with sampling rate half of the
258 // input signal.
259 GetSubframesPitchParameters(
260 kSampleRateHz / 2, gains, lags, kNumPitchSubframes, kNum10msSubframes,
261 &log_old_gain_, &old_lag_, log_pitch_gains, pitch_lags_hz);
262 }
263
Rms(double * rms,size_t length_rms)264 void VadAudioProc::Rms(double* rms, size_t length_rms) {
265 assert(length_rms >= kNum10msSubframes);
266 size_t offset = kNumPastSignalSamples;
267 for (size_t i = 0; i < kNum10msSubframes; i++) {
268 rms[i] = 0;
269 for (size_t n = 0; n < kNumSubframeSamples; n++, offset++)
270 rms[i] += audio_buffer_[offset] * audio_buffer_[offset];
271 rms[i] = sqrt(rms[i] / kNumSubframeSamples);
272 }
273 }
274
275 } // namespace webrtc
276