diff --git a/speechx/examples/feat/streaming_feat_main.cc b/speechx/examples/feat/streaming_feat_main.cc index 29133045497ab4ad1683802710fda0edb403926d..b3ee98422a83f5d8d8282049d677e3f1c0d46a3e 100644 --- a/speechx/examples/feat/streaming_feat_main.cc +++ b/speechx/examples/feat/streaming_feat_main.cc @@ -1,17 +1,34 @@ // todo refactor, repalce with gtest +#include "base/log.h" +#include "base/flags.h" #include "frontend/linear_spectrogram.h" #include "frontend/normalizer.h" #include "frontend/feature_extractor_interface.h" +#include "frontend/raw_audio.h" #include "kaldi/util/table-types.h" -#include "base/log.h" -#include "base/flags.h" #include "kaldi/feat/wave-reader.h" #include "kaldi/util/kaldi-io.h" DEFINE_string(wav_rspecifier, "", "test wav path"); DEFINE_string(feature_wspecifier, "", "test wav ark"); -DEFINE_string(cmvn_path, "./cmvn.ark", "test wav ark"); +DEFINE_string(feature_check_wspecifier, "", "test wav ark"); +DEFINE_string(cmvn_write_path, "./cmvn.ark", "test wav ark"); + + +std::vector mean_{-13730251.531853663, -12982852.199316509, -13673844.299583456, -13089406.559646806, -12673095.524938712, -12823859.223276224, -13590267.158903603, -14257618.467152044, -14374605.116185192, -14490009.21822485, -14849827.158924166, -15354435.470563512, -15834149.206532761, -16172971.985514281, -16348740.496746974, -16423536.699409386, -16556246.263649225, -16744088.772748645, -16916184.08510357, -17054034.840031497, -17165612.509455364, -17255955.470915023, -17322572.527648456, -17408943.862033736, -17521554.799865916, -17620623.254924215, -17699792.395918526, -17723364.411134344, -17741483.4433254, -17747426.888704527, -17733315.928209435, -17748780.160905756, -17808336.883775543, -17895918.671983004, -18009812.59173023, -18098188.66548325, -18195798.958462656, -18293617.62980999, -18397432.92077201, -18505834.787318766, -18585451.8100908, -18652438.235649142, -18700960.306275308, -18734944.58792185, -18737426.313365128, -18735347.165987637, -18738813.444170244, -18737086.848890636, -18731576.2474336, -18717405.44095871, -18703089.25545657, -18691014.546456724, -18692460.568905357, -18702119.628629155, -18727710.621126678, -18761582.72034647, -18806745.835547544, -18850674.8692112, -18884431.510951452, -18919999.992506847, -18939303.799078144, -18952946.273760635, -18980289.22996379, -19011610.17803294, -19040948.61805145, -19061021.429847397, -19112055.53768819, -19149667.414264943, -19201127.05091321, -19270250.82564605, -19334606.883057203, -19390513.336589377, -19444176.259208687, -19502755.000038862, -19544333.014549147, -19612668.183176614, -19681902.19006569, -19771969.951249883, -19873329.723376893, -19996752.59235844, -20110031.131400537, -20231658.612529557, -20319378.894054495, -20378534.45718066, -20413332.089584175, -20438147.844177883, -20443710.248040095, -20465457.02238927, -20488610.969337028, -20516295.16424432, -20541423.795738827, -20553192.874953747, -20573605.50701977, -20577871.61936797, -20571807.008916274, -20556242.38912231, -20542199.30819195, -20521239.063551214, -20519150.80004532, -20527204.80248933, -20536933.769257784, -20543470.522332076, -20549700.089992985, -20551525.24958494, -20554873.406493705, -20564277.65794227, -20572211.740052115, -20574305.69550465, -20575494.450104576, -20567092.577932164, -20549302.929608088, -20545445.11878376, -20546625.326603737, -20549190.03499401, -20554824.947828256, -20568341.378989458, -20577582.331383612, -20577980.519402675, -20566603.03458152, -20560131.592262644, -20552166.469060015, -20549063.06763577, -20544490.562339947, -20539817.82346569, -20528747.715731595, -20518026.24576161, -20510977.844974525, -20506874.36087992, -20506731.11977665, -20510482.133420516, -20507760.92101862, -20494644.834457114, -20480107.89304893, -20461312.091867123, -20442941.75080173, -20426123.02834838, -20424607.675283, -20426810.369107097, -20434024.50097819, -20437404.75544205, -20447688.63916367, -20460893.335563846, -20482922.735127095, -20503610.119434915, -20527062.76448319, -20557830.035128627, -20593274.72068722, -20632528.452965066, -20673637.471334763, -20733106.97143075, -20842921.0447562, -21054357.83621519, -21416569.534189366, -21978460.272811692, -22753170.052172784, -23671344.10563395, -24613499.293358143, -25406477.12230188, -25884377.82156489, -26049040.62791664, -26996879.104431007}; +std::vector variance_{213747175.10846674, 188395815.34302503, 212706429.10966414, 199109025.81461075, 189235901.23864496, 194901336.53253657, 217481594.29306737, 238689869.12327808, 243977501.24115244, 248479623.6431067, 259766741.47116545, 275516766.7790273, 291271202.3691234, 302693239.8220509, 308627358.3997694, 311143911.38788426, 315446105.07731867, 321705430.9341829, 327458907.4659941, 332245072.43223983, 336251717.5935284, 339694069.7639722, 342188204.4322228, 345587110.31313115, 349903086.2875232, 353660214.20643026, 356700344.5270885, 357665362.3529641, 358493352.05658793, 358857951.620328, 358375239.52774596, 358899733.6342954, 361051818.3511561, 364361716.05025816, 368750322.3771452, 372047800.6462831, 375655861.1349018, 379358519.1980013, 383327605.3935181, 387458599.282341, 390434692.3406868, 392994486.35057056, 394874418.04603153, 396230525.79763395, 396365592.0414835, 396334819.8242737, 396488353.19250053, 396438877.00744957, 396197980.4459586, 395590921.6672991, 395001107.62072515, 394528291.7318225, 394593110.424006, 395018405.59353715, 396110577.5415993, 397506704.0371068, 399400197.4657644, 401243568.2468382, 402687134.7805103, 404136047.2872507, 404883170.001883, 405522253.219517, 406660365.3626476, 407919346.0991902, 409045348.5384909, 409759588.7889818, 411974821.8564483, 413489718.78201455, 415535392.56684107, 418466481.97674364, 421104678.35678065, 423405392.5200779, 425550570.40798235, 427929423.9579701, 429585274.253478, 432368493.55181056, 435193587.13513297, 438886855.20476013, 443058876.8633751, 448181232.5093362, 452883835.6332396, 458056721.77926534, 461816531.22735566, 464363620.1970998, 465886343.5057493, 466928872.0651, 467180536.42647296, 468111848.70714295, 469138695.3071312, 470378429.6930793, 471517958.7132626, 472109050.4262365, 473087417.0177867, 473381322.04648733, 473220195.85483915, 472666071.8998819, 472124669.87879956, 471298571.411737, 471251033.2902761, 471672676.43128747, 472177147.2193172, 472572361.7711908, 472968783.7751127, 473156295.4164052, 473398034.82676554, 473897703.5203811, 474328271.33112127, 474452670.98002136, 474549003.99284613, 474252887.13567275, 473557462.909069, 473483385.85193115, 473609738.04855174, 473746944.82085115, 474016729.91696435, 474617321.94138587, 475045097.237122, 475125402.586558, 474664112.9824912, 474426247.5800283, 474104075.42796475, 473978219.7273978, 473773171.7798875, 473578534.69508696, 473102924.16904145, 472651240.5232615, 472374383.1810912, 472209479.6956096, 472202298.8921673, 472370090.76781124, 472220933.99374026, 471625467.37106377, 470994646.51883453, 470182428.9637543, 469348211.5939578, 468570387.4467277, 468540442.7225135, 468672018.90414184, 468994346.9533251, 469138757.58201426, 469553915.95710236, 470134523.38582784, 471082421.62055486, 471962316.51804745, 472939745.1708408, 474250621.5944825, 475773933.43199486, 477465399.71087736, 479218782.61382693, 481752299.7930922, 486608947.8984568, 496119403.2067917, 512730085.5704984, 539048915.2641417, 576285298.3548826, 621610270.2240586, 669308196.4436442, 710656993.5957186, 736344437.3725077, 745481288.0241544, 801121432.9925804}; +int count_ = 912592; + +void WriteMatrix() { + kaldi::Matrix cmvn_stats(2, mean_.size()+ 1); + for (size_t idx = 0; idx < mean_.size(); ++idx) { + cmvn_stats(0, idx) = mean_[idx]; + cmvn_stats(1, idx) = variance_[idx]; + } + cmvn_stats(0, mean_.size()) = count_; + kaldi::WriteKaldiObject(cmvn_stats, FLAGS_cmvn_write_path, true); +} int main(int argc, char* argv[]) { gflags::ParseCommandLineFlags(&argc, &argv, false); @@ -19,33 +36,69 @@ int main(int argc, char* argv[]) { kaldi::SequentialTableReader wav_reader(FLAGS_wav_rspecifier); kaldi::BaseFloatMatrixWriter feat_writer(FLAGS_feature_wspecifier); + WriteMatrix(); // test feature linear_spectorgram: wave --> decibel_normalizer --> hanning window -->linear_spectrogram --> cmvn - // --> feature_cache int32 num_done = 0, num_err = 0; + std::unique_ptr data_source(new ppspeech::RawDataSource()); + ppspeech::LinearSpectrogramOptions opt; opt.frame_opts.frame_length_ms = 20; opt.frame_opts.frame_shift_ms = 10; ppspeech::DecibelNormalizerOptions db_norm_opt; std::unique_ptr base_feature_extractor( - new ppspeech::DecibelNormalizer(db_norm_opt)); + new ppspeech::DecibelNormalizer(db_norm_opt, std::move(data_source))); - std::shared_ptr linear_spectrogram( - new ppspeech::LinearSpectrogram(opt, base_feature_extractor)); + std::unique_ptr linear_spectrogram( + new ppspeech::LinearSpectrogram(opt, std::move(base_feature_extractor))); - std::shared_ptr cmvn( - new ppspeech::CMVN(FLAGS_cmvn_path, linear_spectrogram); - ppspeech::FeatureCache(cmvn); + ppspeech::CMVN cmvn(FLAGS_cmvn_write_path, std::move(linear_spectrogram)); float streaming_chunk = 0.36; int sample_rate = 16000; int chunk_sample_size = streaming_chunk * sample_rate; - // thread 1 feed feature for (; !wav_reader.Done(); wav_reader.Next()) { std::string utt = wav_reader.Key(); const kaldi::WaveData &wave_data = wav_reader.Value(); + int32 this_channel = 0; + kaldi::SubVector waveform(wave_data.Data(), this_channel); + int tot_samples = waveform.Dim(); + int sample_offset = 0; + std::vector> feats; + int feature_rows = 0; + while (sample_offset < tot_samples) { + int cur_chunk_size = std::min(chunk_sample_size, tot_samples - sample_offset); + kaldi::Vector wav_chunk(cur_chunk_size); + for (int i = 0; i < cur_chunk_size; ++i) { + wav_chunk(i) = waveform(sample_offset + i); + } + kaldi::Vector features; + cmvn.AcceptWaveform(wav_chunk); + cmvn.Read(&features); + + std::cout << wav_chunk(0) << std::endl; + std::cout << features(0) << std::endl; + + feats.push_back(features); + sample_offset += cur_chunk_size; + feature_rows += features.Dim() / cmvn.Dim(); + } + + int cur_idx = 0; + kaldi::Matrix features(feature_rows, cmvn.Dim()); + for (auto feat : feats) { + int num_rows = feat.Dim() / cmvn.Dim(); + for (int row_idx = 0; row_idx < num_rows; ++row_idx) { + for (int col_idx = 0; col_idx < cmvn.Dim(); ++col_idx) { + features(cur_idx, col_idx) = feat(row_idx*cmvn.Dim() + col_idx); + } + ++cur_idx; + } + } + feat_writer.Write(utt, features); + if (num_done % 50 == 0 && num_done != 0) KALDI_VLOG(2) << "Processed " << num_done << " utterances"; num_done++; diff --git a/speechx/speechx/base/common.h b/speechx/speechx/base/common.h index 3b58f73cd9da77eab03f2a13d4e41241faa1b8b8..ac01a9778be1ffbbcebc0c668be9c61d0d55f7e0 100644 --- a/speechx/speechx/base/common.h +++ b/speechx/speechx/base/common.h @@ -15,22 +15,23 @@ #pragma once #include -#include #include #include +#include #include #include -#include #include #include #include #include #include +#include #include #include -#include +#include +#include -#include "base/basic_types.h" -#include "base/flags.h" #include "base/log.h" +#include "base/flags.h" +#include "base/basic_types.h" #include "base/macros.h" diff --git a/speechx/speechx/frontend/CMakeLists.txt b/speechx/speechx/frontend/CMakeLists.txt index da81a481bbe4fbb63f3ea6ff4c1f2fb46c3125ec..e43bd182f1d78c758d5ab84b6ce61b96181535a8 100644 --- a/speechx/speechx/frontend/CMakeLists.txt +++ b/speechx/speechx/frontend/CMakeLists.txt @@ -2,7 +2,8 @@ project(frontend) add_library(frontend STATIC normalizer.cc - linear_spectrogram.cc + linear_spectrogram.cc + raw_audio.cc ) target_link_libraries(frontend PUBLIC kaldi-matrix) \ No newline at end of file diff --git a/speechx/speechx/frontend/feature_extractor_interface.h b/speechx/speechx/frontend/feature_extractor_interface.h index e39f5e465458b619d39be1a22e7c3d97131a621a..fc06f24af88ad9b23614a8bc2f4797be15caf4c1 100644 --- a/speechx/speechx/frontend/feature_extractor_interface.h +++ b/speechx/speechx/frontend/feature_extractor_interface.h @@ -21,9 +21,8 @@ namespace ppspeech { class FeatureExtractorInterface { public: - virtual void AcceptWaveform( - const kaldi::VectorBase& input) = 0; - virtual void Read(kaldi::VectorBase* feat) = 0; + virtual void AcceptWaveform(const kaldi::VectorBase& input) = 0; + virtual void Read(kaldi::Vector* feat) = 0; virtual size_t Dim() const = 0; }; diff --git a/speechx/speechx/frontend/linear_spectrogram.cc b/speechx/speechx/frontend/linear_spectrogram.cc index 6c008c399cc357f8a00050dd1b3a294a597bdd87..ed4c2977e39f5be67b7581b60bce8c1185adc276 100644 --- a/speechx/speechx/frontend/linear_spectrogram.cc +++ b/speechx/speechx/frontend/linear_spectrogram.cc @@ -25,153 +25,146 @@ using kaldi::VectorBase; using kaldi::Matrix; using std::vector; -// todo remove later +//todo remove later void CopyVector2StdVector_(const VectorBase& input, - vector* output) { - if (input.Dim() == 0) return; - output->resize(input.Dim()); - for (size_t idx = 0; idx < input.Dim(); ++idx) { - (*output)[idx] = input(idx); - } + vector* output) { + if (input.Dim() == 0) return; + output->resize(input.Dim()); + for (size_t idx = 0; idx < input.Dim(); ++idx) { + (*output)[idx] = input(idx); + } } void CopyStdVector2Vector_(const vector& input, - Vector* output) { - if (input.empty()) return; - output->Resize(input.size()); - for (size_t idx = 0; idx < input.size(); ++idx) { - (*output)(idx) = input[idx]; - } + Vector* output) { + if (input.empty()) return; + output->Resize(input.size()); + for (size_t idx = 0; idx < input.size(); ++idx) { + (*output)(idx) = input[idx]; + } } LinearSpectrogram::LinearSpectrogram( const LinearSpectrogramOptions& opts, std::unique_ptr base_extractor) { - opts_ = opts; - base_extractor_ = std::move(base_extractor); - int32 window_size = opts.frame_opts.WindowSize(); - int32 window_shift = opts.frame_opts.WindowShift(); - fft_points_ = window_size; - hanning_window_.resize(window_size); - - double a = M_2PI / (window_size - 1); - hanning_window_energy_ = 0; - for (int i = 0; i < window_size; ++i) { - hanning_window_[i] = 0.5 - 0.5 * cos(a * i); - hanning_window_energy_ += hanning_window_[i] * hanning_window_[i]; - } - - dim_ = fft_points_ / 2 + 1; // the dimension is Fs/2 Hz + opts_ = opts; + base_extractor_ = std::move(base_extractor); + int32 window_size = opts.frame_opts.WindowSize(); + int32 window_shift = opts.frame_opts.WindowShift(); + fft_points_ = window_size; + chunk_sample_size_ = + static_cast(opts.streaming_chunk * opts.frame_opts.samp_freq); + hanning_window_.resize(window_size); + + double a = M_2PI / (window_size - 1); + hanning_window_energy_ = 0; + for (int i = 0; i < window_size; ++i) { + hanning_window_[i] = 0.5 - 0.5 * cos(a * i); + hanning_window_energy_ += hanning_window_[i] * hanning_window_[i]; + } + + dim_ = fft_points_ / 2 + 1; // the dimension is Fs/2 Hz } void LinearSpectrogram::AcceptWaveform(const VectorBase& input) { base_extractor_->AcceptWaveform(input); } +void LinearSpectrogram::Read(Vector* feat) { + Vector input_feats(chunk_sample_size_); + base_extractor_->Read(&input_feats); + vector input_feats_vec(input_feats.Dim()); + CopyVector2StdVector_(input_feats, &input_feats_vec); + //for (int idx = 0; idx < input_feats.Dim(); ++idx) { + // input_feats_vec[idx] = input_feats(idx); + //} + vector> result; + Compute(input_feats_vec, result); + int32 feat_size = 0; + if (result.size() != 0) { + feat_size = result.size() * result[0].size(); + } + feat->Resize(feat_size); + for (size_t idx = 0; idx < feat_size; ++idx) { + (*feat)(idx) = result[idx / dim_][idx % dim_]; + } + return; +} + void LinearSpectrogram::Hanning(vector* data) const { - CHECK_GE(data->size(), hanning_window_.size()); + CHECK_GE(data->size(), hanning_window_.size()); - for (size_t i = 0; i < hanning_window_.size(); ++i) { - data->at(i) *= hanning_window_[i]; - } + for (size_t i = 0; i < hanning_window_.size(); ++i) { + data->at(i) *= hanning_window_[i]; + } } bool LinearSpectrogram::NumpyFft(vector* v, vector* real, vector* img) const { - Vector v_tmp; - CopyStdVector2Vector_(*v, &v_tmp); - RealFft(&v_tmp, true); - CopyVector2StdVector_(v_tmp, v); - real->push_back(v->at(0)); - img->push_back(0); - for (int i = 1; i < v->size() / 2; i++) { - real->push_back(v->at(2 * i)); - img->push_back(v->at(2 * i + 1)); - } - real->push_back(v->at(1)); - img->push_back(0); - - return true; -} - -// todo remove later -void LinearSpectrogram::ReadFeats(Matrix* feats) { - Vector tmp; - waveform_.Resize(base_extractor_->Dim()); - Compute(tmp, &waveform_); - vector> result; - vector feats_vec; - CopyVector2StdVector_(waveform_, &feats_vec); - Compute(feats_vec, result); - feats->Resize(result.size(), result[0].size()); - for (int row_idx = 0; row_idx < result.size(); ++row_idx) { - for (int col_idx = 0; col_idx < result[0].size(); ++col_idx) { - (*feats)(row_idx, col_idx) = result[row_idx][col_idx]; - } - } - waveform_.Resize(0); -} - -void LinearSpectrogram::Read(VectorBase* feat) { - // todo - return; -} - -// only for test, remove later -// todo: compute the feature frame by frame. -void LinearSpectrogram::Compute(const VectorBase& input, - VectorBase* feature) { - base_extractor_->Read(feature); + Vector v_tmp; + CopyStdVector2Vector_(*v, &v_tmp); + RealFft(&v_tmp, true); + CopyVector2StdVector_(v_tmp, v); + real->push_back(v->at(0)); + img->push_back(0); + for (int i = 1; i < v->size() / 2; i++) { + real->push_back(v->at(2 * i)); + img->push_back(v->at(2 * i + 1)); + } + real->push_back(v->at(1)); + img->push_back(0); + + return true; } // Compute spectrogram feat, only for test, remove later // todo: refactor later (SmileGoat) bool LinearSpectrogram::Compute(const vector& wave, vector>& feat) { - int num_samples = wave.size(); - const int& frame_length = opts_.frame_opts.WindowSize(); - const int& sample_rate = opts_.frame_opts.samp_freq; - const int& frame_shift = opts_.frame_opts.WindowShift(); - const int& fft_points = fft_points_; - const float scale = hanning_window_energy_ * sample_rate; - - if (num_samples < frame_length) { - return true; - } - - int num_frames = 1 + ((num_samples - frame_length) / frame_shift); - feat.resize(num_frames); - vector fft_real((fft_points_ / 2 + 1), 0); - vector fft_img((fft_points_ / 2 + 1), 0); - vector v(frame_length, 0); - vector power((fft_points / 2 + 1)); - - for (int i = 0; i < num_frames; ++i) { - vector data(wave.data() + i * frame_shift, - wave.data() + i * frame_shift + frame_length); - Hanning(&data); - fft_img.clear(); - fft_real.clear(); - v.assign(data.begin(), data.end()); - NumpyFft(&v, &fft_real, &fft_img); - - feat[i].resize(fft_points / 2 + 1); // the last dimension is Fs/2 Hz - for (int j = 0; j < (fft_points / 2 + 1); ++j) { - power[j] = fft_real[j] * fft_real[j] + fft_img[j] * fft_img[j]; - feat[i][j] = power[j]; - - if (j == 0 || j == feat[0].size() - 1) { - feat[i][j] /= scale; - } else { - feat[i][j] *= (2.0 / scale); - } - - // log added eps=1e-14 - feat[i][j] = std::log(feat[i][j] + 1e-14); - } + int num_samples = wave.size(); + const int& frame_length = opts_.frame_opts.WindowSize(); + const int& sample_rate = opts_.frame_opts.samp_freq; + const int& frame_shift = opts_.frame_opts.WindowShift(); + const int& fft_points = fft_points_; + const float scale = hanning_window_energy_ * sample_rate; + + if (num_samples < frame_length) { + return true; + } + + int num_frames = 1 + ((num_samples - frame_length) / frame_shift); + feat.resize(num_frames); + vector fft_real((fft_points_ / 2 + 1), 0); + vector fft_img((fft_points_ / 2 + 1), 0); + vector v(frame_length, 0); + vector power((fft_points / 2 + 1)); + + for (int i = 0; i < num_frames; ++i) { + vector data(wave.data() + i * frame_shift, + wave.data() + i * frame_shift + frame_length); + Hanning(&data); + fft_img.clear(); + fft_real.clear(); + v.assign(data.begin(), data.end()); + NumpyFft(&v, &fft_real, &fft_img); + + feat[i].resize(fft_points / 2 + 1); // the last dimension is Fs/2 Hz + for (int j = 0; j < (fft_points / 2 + 1); ++j) { + power[j] = fft_real[j] * fft_real[j] + fft_img[j] * fft_img[j]; + feat[i][j] = power[j]; + + if (j == 0 || j == feat[0].size() - 1) { + feat[i][j] /= scale; + } else { + feat[i][j] *= (2.0 / scale); + } + + // log added eps=1e-14 + feat[i][j] = std::log(feat[i][j] + 1e-14); } - return true; + } + return true; } } // namespace ppspeech \ No newline at end of file diff --git a/speechx/speechx/frontend/linear_spectrogram.h b/speechx/speechx/frontend/linear_spectrogram.h index 20b5e4b58274786899f8dce30edcfc21376075e0..e4dc3e33d138d57f683ce583ad625285822a0ab7 100644 --- a/speechx/speechx/frontend/linear_spectrogram.h +++ b/speechx/speechx/frontend/linear_spectrogram.h @@ -1,45 +1,35 @@ -// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - #pragma once -#include "base/common.h" #include "frontend/feature_extractor_interface.h" #include "kaldi/feat/feature-window.h" +#include "base/common.h" namespace ppspeech { struct LinearSpectrogramOptions { kaldi::FrameExtractionOptions frame_opts; - LinearSpectrogramOptions() : frame_opts() {} - - void Register(kaldi::OptionsItf* opts) { frame_opts.Register(opts); } + kaldi::BaseFloat streaming_chunk; + LinearSpectrogramOptions(): + streaming_chunk(0.36), + frame_opts() {} + + void Register(kaldi::OptionsItf* opts) { + opts->Register("streaming-chunk", &streaming_chunk, "streaming chunk size"); + frame_opts.Register(opts); + } }; class LinearSpectrogram : public FeatureExtractorInterface { public: - explicit LinearSpectrogram( - const LinearSpectrogramOptions& opts, - std::unique_ptr base_extractor); - virtual void AcceptWaveform( - const kaldi::VectorBase& input); - virtual void Read(kaldi::VectorBase* feat); + explicit LinearSpectrogram(const LinearSpectrogramOptions& opts, + std::unique_ptr base_extractor); + virtual void AcceptWaveform(const kaldi::VectorBase& input); + virtual void Read(kaldi::Vector* feat); virtual size_t Dim() const { return dim_; } void ReadFeats(kaldi::Matrix* feats); - private: + private: void Hanning(std::vector* data) const; bool Compute(const std::vector& wave, std::vector>& feat); @@ -54,8 +44,9 @@ class LinearSpectrogram : public FeatureExtractorInterface { std::vector hanning_window_; kaldi::BaseFloat hanning_window_energy_; LinearSpectrogramOptions opts_; - kaldi::Vector waveform_; // remove later, todo(SmileGoat) + kaldi::Vector waveform_; // remove later, todo(SmileGoat) std::unique_ptr base_extractor_; + int chunk_sample_size_; DISALLOW_COPY_AND_ASSIGN(LinearSpectrogram); }; diff --git a/speechx/speechx/frontend/normalizer.cc b/speechx/speechx/frontend/normalizer.cc index abf798e5787d9fd75c3c288e29245312a226526a..69c9ab5934c9cd5286837d85ee459c8c0ceb6cf1 100644 --- a/speechx/speechx/frontend/normalizer.cc +++ b/speechx/speechx/frontend/normalizer.cc @@ -1,17 +1,3 @@ -// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - #include "frontend/normalizer.h" #include "kaldi/feat/cmvn.h" @@ -24,175 +10,176 @@ using kaldi::VectorBase; using kaldi::BaseFloat; using std::vector; using kaldi::SubVector; +using std::unique_ptr; -DecibelNormalizer::DecibelNormalizer(const DecibelNormalizerOptions& opts) { - opts_ = opts; - dim_ = 0; +DecibelNormalizer::DecibelNormalizer(const DecibelNormalizerOptions& opts, + std::unique_ptr base_extractor) { + base_extractor_ = std::move(base_extractor); + opts_ = opts; + dim_ = 0; } - -void DecibelNormalizer::AcceptWaveform( - const kaldi::VectorBase& input) { - dim_ = input.Dim(); - waveform_.Resize(input.Dim()); - waveform_.CopyFromVec(input); + +void DecibelNormalizer::AcceptWaveform(const kaldi::VectorBase& input) { + //dim_ = input.Dim(); + //waveform_.Resize(input.Dim()); + //waveform_.CopyFromVec(input); + base_extractor_->AcceptWaveform(input); } -void DecibelNormalizer::Read(kaldi::VectorBase* feat) { - if (waveform_.Dim() == 0) return; - Compute(waveform_, feat); +void DecibelNormalizer::Read(kaldi::Vector* feat) { + // if (waveform_.Dim() == 0) return; + base_extractor_->Read(feat); + Compute(feat); } -// todo remove later +//todo remove later void CopyVector2StdVector(const kaldi::VectorBase& input, vector* output) { - if (input.Dim() == 0) return; - output->resize(input.Dim()); - for (size_t idx = 0; idx < input.Dim(); ++idx) { - (*output)[idx] = input(idx); - } + if (input.Dim() == 0) return; + output->resize(input.Dim()); + for (size_t idx = 0; idx < input.Dim(); ++idx) { + (*output)[idx] = input(idx); + } } void CopyStdVector2Vector(const vector& input, VectorBase* output) { - if (input.empty()) return; - assert(input.size() == output->Dim()); - for (size_t idx = 0; idx < input.size(); ++idx) { - (*output)(idx) = input[idx]; - } + if (input.empty()) return; + assert(input.size() == output->Dim()); + for (size_t idx = 0; idx < input.size(); ++idx) { + (*output)(idx) = input[idx]; + } } -bool DecibelNormalizer::Compute(const VectorBase& input, - VectorBase* feat) const { - // calculate db rms - BaseFloat rms_db = 0.0; - BaseFloat mean_square = 0.0; - BaseFloat gain = 0.0; - BaseFloat wave_float_normlization = 1.0f / (std::pow(2, 16 - 1)); - - vector samples; - samples.resize(input.Dim()); - for (int32 i = 0; i < samples.size(); ++i) { - samples[i] = input(i); - } - - // square - for (auto& d : samples) { - if (opts_.convert_int_float) { - d = d * wave_float_normlization; - } - mean_square += d * d; +bool DecibelNormalizer::Compute(VectorBase* feat) const { + // calculate db rms + BaseFloat rms_db = 0.0; + BaseFloat mean_square = 0.0; + BaseFloat gain = 0.0; + BaseFloat wave_float_normlization = 1.0f / (std::pow(2, 16 - 1)); + + vector samples; + samples.resize(feat->Dim()); + for (size_t i = 0; i < samples.size(); ++i) { + samples[i] = (*feat)(i); + } + + // square + for (auto &d : samples) { + if (opts_.convert_int_float) { + d = d * wave_float_normlization; } - - // mean - mean_square /= samples.size(); - rms_db = 10 * std::log10(mean_square); - gain = opts_.target_db - rms_db; - - if (gain > opts_.max_gain_db) { - LOG(ERROR) - << "Unable to normalize segment to " << opts_.target_db << "dB," - << "because the the probable gain have exceeds opts_.max_gain_db" - << opts_.max_gain_db << "dB."; - return false; - } - - // Note that this is an in-place transformation. - for (auto& item : samples) { - // python item *= 10.0 ** (gain / 20.0) - item *= std::pow(10.0, gain / 20.0); - } - - CopyStdVector2Vector(samples, feat); - return true; + mean_square += d * d; + } + + // mean + mean_square /= samples.size(); + rms_db = 10 * std::log10(mean_square); + gain = opts_.target_db - rms_db; + + if (gain > opts_.max_gain_db) { + LOG(ERROR) << "Unable to normalize segment to " << opts_.target_db << "dB," + << "because the the probable gain have exceeds opts_.max_gain_db" + << opts_.max_gain_db << "dB."; + return false; + } + + // Note that this is an in-place transformation. + for (auto &item : samples) { + // python item *= 10.0 ** (gain / 20.0) + item *= std::pow(10.0, gain / 20.0); + } + + CopyStdVector2Vector(samples, feat); + return true; } -CMVN::CMVN(std::string cmvn_file) : var_norm_(true) { +CMVN::CMVN(std::string cmvn_file, + unique_ptr base_extractor) + : var_norm_(true) { + base_extractor_ = std::move(base_extractor); bool binary; kaldi::Input ki(cmvn_file, &binary); stats_.Read(ki.Stream(), binary); + dim_ = stats_.NumCols() - 1; } void CMVN::AcceptWaveform(const kaldi::VectorBase& input) { + base_extractor_->AcceptWaveform(input); return; } -void CMVN::Read(kaldi::VectorBase* feat) { return; } +void CMVN::Read(kaldi::Vector* feat) { + base_extractor_->Read(feat); + Compute(feat); + return; +} // feats contain num_frames feature. -void CMVN::ApplyCMVN(bool var_norm, VectorBase* feats) { - KALDI_ASSERT(feats != NULL); - int32 dim = stats_.NumCols() - 1; - if (stats_.NumRows() > 2 || stats_.NumRows() < 1 || - feats->Dim() % dim != 0) { - KALDI_ERR << "Dim mismatch: cmvn " << stats_.NumRows() << 'x' - << stats_.NumCols() << ", feats " << feats->Dim() << 'x'; +void CMVN::Compute(VectorBase* feats) const { + KALDI_ASSERT(feats != NULL); + int32 dim = stats_.NumCols() - 1; + if (stats_.NumRows() > 2 || stats_.NumRows() < 1 || feats->Dim() % dim != 0) { + KALDI_ERR << "Dim mismatch: cmvn " + << stats_.NumRows() << 'x' << stats_.NumCols() + << ", feats " << feats->Dim() << 'x'; + } + if (stats_.NumRows() == 1 && var_norm_) { + KALDI_ERR << "You requested variance normalization but no variance stats_ " + << "are supplied."; + } + + double count = stats_(0, dim); + // Do not change the threshold of 1.0 here: in the balanced-cmvn code, when + // computing an offset and representing it as stats_, we use a count of one. + if (count < 1.0) + KALDI_ERR << "Insufficient stats_ for cepstral mean and variance normalization: " + << "count = " << count; + + if (!var_norm_) { + Vector offset(feats->Dim()); + SubVector mean_stats(stats_.RowData(0), dim); + Vector mean_stats_apply(feats->Dim()); + //fill the datat of mean_stats in mean_stats_appy whose dim is equal with the dim of feature. + //the dim of feats = dim * num_frames; + for (int32 idx = 0; idx < feats->Dim() / dim; ++idx) { + SubVector stats_tmp(mean_stats_apply.Data() + dim*idx, dim); + stats_tmp.CopyFromVec(mean_stats); } - if (stats_.NumRows() == 1 && var_norm) { - KALDI_ERR - << "You requested variance normalization but no variance stats_ " - << "are supplied."; - } - - double count = stats_(0, dim); - // Do not change the threshold of 1.0 here: in the balanced-cmvn code, when - // computing an offset and representing it as stats_, we use a count of one. - if (count < 1.0) - KALDI_ERR << "Insufficient stats_ for cepstral mean and variance " - "normalization: " - << "count = " << count; - - if (!var_norm) { - Vector offset(feats->Dim()); - SubVector mean_stats(stats_.RowData(0), dim); - Vector mean_stats_apply(feats->Dim()); - // fill the datat of mean_stats in mean_stats_appy whose dim is equal - // with the dim of feature. - // the dim of feats = dim * num_frames; - for (int32 idx = 0; idx < feats->Dim() / dim; ++idx) { - SubVector stats_tmp(mean_stats_apply.Data() + dim * idx, - dim); - stats_tmp.CopyFromVec(mean_stats); - } - offset.AddVec(-1.0 / count, mean_stats_apply); - feats->AddVec(1.0, offset); - return; + offset.AddVec(-1.0 / count, mean_stats_apply); + feats->AddVec(1.0, offset); + return; + } + // norm(0, d) = mean offset; + // norm(1, d) = scale, e.g. x(d) <-- x(d)*norm(1, d) + norm(0, d). + kaldi::Matrix norm(2, feats->Dim()); + for (int32 d = 0; d < dim; d++) { + double mean, offset, scale; + mean = stats_(0, d)/count; + double var = (stats_(1, d)/count) - mean*mean, + floor = 1.0e-20; + if (var < floor) { + KALDI_WARN << "Flooring cepstral variance from " << var << " to " + << floor; + var = floor; } - // norm(0, d) = mean offset; - // norm(1, d) = scale, e.g. x(d) <-- x(d)*norm(1, d) + norm(0, d). - kaldi::Matrix norm(2, feats->Dim()); - for (int32 d = 0; d < dim; d++) { - double mean, offset, scale; - mean = stats_(0, d) / count; - double var = (stats_(1, d) / count) - mean * mean, floor = 1.0e-20; - if (var < floor) { - KALDI_WARN << "Flooring cepstral variance from " << var << " to " - << floor; - var = floor; - } - scale = 1.0 / sqrt(var); - if (scale != scale || 1 / scale == 0.0) - KALDI_ERR - << "NaN or infinity in cepstral mean/variance computation"; - offset = -(mean * scale); - for (int32 d_skip = d; d_skip < feats->Dim();) { - norm(0, d_skip) = offset; - norm(1, d_skip) = scale; - d_skip = d_skip + dim; - } + scale = 1.0 / sqrt(var); + if (scale != scale || 1/scale == 0.0) + KALDI_ERR << "NaN or infinity in cepstral mean/variance computation"; + offset = -(mean*scale); + for (int32 d_skip = d; d_skip < feats->Dim();) { + norm(0, d_skip) = offset; + norm(1, d_skip) = scale; + d_skip = d_skip + dim; } - // Apply the normalization. - feats->MulElements(norm.Row(1)); - feats->AddVec(1.0, norm.Row(0)); + } + // Apply the normalization. + feats->MulElements(norm.Row(1)); + feats->AddVec(1.0, norm.Row(0)); } -void CMVN::ApplyCMVNMatrix(bool var_norm, kaldi::MatrixBase* feats) { - ApplyCmvn(stats_, var_norm, feats); +void CMVN::ApplyCMVN(kaldi::MatrixBase* feats) { + ApplyCmvn(stats_, var_norm_, feats); } -bool CMVN::Compute(const VectorBase& input, - VectorBase* feat) const { - return false; -} - - -} // namespace ppspeech +} // namespace ppspeech diff --git a/speechx/speechx/frontend/normalizer.h b/speechx/speechx/frontend/normalizer.h index 6af5cdd8fd729f9ad029795ad42bc1bca4bf82d3..13c5b8df9ad910f3f606343ad7e9546fd4b37e0a 100644 --- a/speechx/speechx/frontend/normalizer.h +++ b/speechx/speechx/frontend/normalizer.h @@ -1,56 +1,40 @@ -// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - #pragma once #include "base/common.h" #include "frontend/feature_extractor_interface.h" -#include "kaldi/matrix/kaldi-matrix.h" #include "kaldi/util/options-itf.h" +#include "kaldi/matrix/kaldi-matrix.h" namespace ppspeech { struct DecibelNormalizerOptions { - float target_db; - float max_gain_db; - bool convert_int_float; - DecibelNormalizerOptions() - : target_db(-20), max_gain_db(300.0), convert_int_float(false) {} + float target_db; + float max_gain_db; + bool convert_int_float; + DecibelNormalizerOptions() : + target_db(-20), + max_gain_db(300.0), + convert_int_float(false){} void Register(kaldi::OptionsItf* opts) { - opts->Register( - "target-db", &target_db, "target db for db normalization"); - opts->Register( - "max-gain-db", &max_gain_db, "max gain db for db normalization"); - opts->Register("convert-int-float", - &convert_int_float, - "if convert int samples to float"); + opts->Register("target-db", &target_db, "target db for db normalization"); + opts->Register("max-gain-db", &max_gain_db, "max gain db for db normalization"); + opts->Register("convert-int-float", &convert_int_float, "if convert int samples to float"); } }; class DecibelNormalizer : public FeatureExtractorInterface { public: - explicit DecibelNormalizer(const DecibelNormalizerOptions& opts); - virtual void AcceptWaveform( - const kaldi::VectorBase& input); - virtual void Read(kaldi::VectorBase* feat); + explicit DecibelNormalizer( + const DecibelNormalizerOptions& opts, + std::unique_ptr base_extractor); + virtual void AcceptWaveform(const kaldi::VectorBase& input); + virtual void Read(kaldi::Vector* feat); virtual size_t Dim() const { return dim_; } - bool Compute(const kaldi::VectorBase& input, - kaldi::VectorBase* feat) const; private: + bool Compute(kaldi::VectorBase* feat) const; DecibelNormalizerOptions opts_; size_t dim_; std::unique_ptr base_extractor_; @@ -60,20 +44,18 @@ class DecibelNormalizer : public FeatureExtractorInterface { class CMVN : public FeatureExtractorInterface { public: - explicit CMVN(std::string cmvn_file); - virtual void AcceptWaveform( - const kaldi::VectorBase& input); - virtual void Read(kaldi::VectorBase* feat); - virtual size_t Dim() const { return stats_.NumCols() - 1; } - bool Compute(const kaldi::VectorBase& input, - kaldi::VectorBase* feat) const; - // for test - void ApplyCMVN(bool var_norm, kaldi::VectorBase* feats); - void ApplyCMVNMatrix(bool var_norm, kaldi::MatrixBase* feats); + explicit CMVN( + std::string cmvn_file, + std::unique_ptr base_extractor); + virtual void AcceptWaveform(const kaldi::VectorBase& input); + virtual void Read(kaldi::Vector* feat); + virtual size_t Dim() const { return dim_; } private: + void Compute(kaldi::VectorBase* feat) const; + void ApplyCMVN(kaldi::MatrixBase* feats); kaldi::Matrix stats_; - std::shared_ptr base_extractor_; + std::unique_ptr base_extractor_; size_t dim_; bool var_norm_; };