| author | Dan Povey <dpovey@gmail.com> | |
| Fri, 8 Aug 2014 22:33:00 +0000 (22:33 +0000) | ||
| committer | Dan Povey <dpovey@gmail.com> | |
| Fri, 8 Aug 2014 22:33:00 +0000 (22:33 +0000) |
git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/online@4284 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8
15 files changed:
diff --cc src/decoder/lattice-faster-decoder.cc
index 1062ef5fbcca7deca3d19e9f27709121adf5c002,ecb4d1642a635379558941706ae6e85f5673254a..294abd52c3ef22cf1e0e6dee9239d02dab9dfd4f
// PruneForwardLinksFinal is a version of PruneForwardLinks that we call
// on the final frame. If there are final tokens active, it uses
// the final-probs for pruning, otherwise it treats all tokens as final.
-void LatticeFasterDecoder::PruneForwardLinksFinal(int32 frame) {
- KALDI_ASSERT(static_cast<size_t>(frame+1) == active_toks_.size());
- if (active_toks_[frame].toks == NULL ) // empty list; should not happen.
+void LatticeFasterDecoder::PruneForwardLinksFinal() {
+ KALDI_ASSERT(!active_toks_.empty());
+ int32 frame_plus_one = active_toks_.size() - 1;
+
+ if (active_toks_[frame_plus_one].toks == NULL ) // empty list; should not happen.
- KALDI_WARN << "No tokens alive at end of file\n";
+ KALDI_WARN << "No tokens alive at end of file";
- // First go through, working out the best token (do it in parallel
- // including final-probs and not including final-probs; we'll take
- // the one with final-probs if it's valid).
- const BaseFloat infinity = std::numeric_limits<BaseFloat>::infinity();
- BaseFloat best_cost_final = infinity,
- best_cost_nofinal = infinity;
- unordered_map<Token*, BaseFloat> tok_to_final_cost;
-
- Elem *cur_toks = toks_.Clear(); // swapping prev_toks_ / cur_toks_
- for (Elem *e = cur_toks; e != NULL; e = e->tail) {
- StateId state = e->key;
- Token *tok = e->val;
- BaseFloat final_cost = fst_.Final(state).Value();
- best_cost_final = std::min(best_cost_final, tok->tot_cost + final_cost);
- tok_to_final_cost[tok] = final_cost;
- best_cost_nofinal = std::min(best_cost_nofinal, tok->tot_cost);
- }
- DeleteElems(cur_toks);
- final_active_ = (best_cost_final != infinity);
-
- // Now go through tokens on this frame, pruning forward links... may have
- // to iterate a few times until there is no more change, because the list is
- // not in topological order.
+ typedef unordered_map<Token*, BaseFloat>::const_iterator IterType;
+ ComputeFinalCosts(&final_costs_, &final_relative_cost_, &final_best_cost_);
+ decoding_finalized_ = true;
+ // We call DeleteElems() as a nicety, not because it's really necessary;
+ // otherwise there would be a time, after calling PruneTokensForFrame() on the
+ // final frame, when toks_.GetList() or toks_.Clear() would contain pointers
+ // to nonexistent tokens.
+ DeleteElems(toks_.Clear());
+ // Now go through tokens on this frame, pruning forward links... may have to
+ // iterate a few times until there is no more change, because the list is not
+ // in topological order. This is a modified version of the code in
+ // PruneForwardLinks, but here we also take account of the final-probs.
bool changed = true;
BaseFloat delta = 1.0e-05;
while (changed) {
// [we don't do this in PruneForwardLinks because it would give us
// a problem with dangling pointers].
// It's called by PruneActiveTokens if any forward links have been pruned
-void LatticeFasterDecoder::PruneTokensForFrame(int32 frame) {
- KALDI_ASSERT(frame >= 0 && frame < active_toks_.size());
- Token *&toks = active_toks_[frame].toks;
+void LatticeFasterDecoder::PruneTokensForFrame(int32 frame_plus_one) {
+ KALDI_ASSERT(frame_plus_one >= 0 && frame_plus_one < active_toks_.size());
+ Token *&toks = active_toks_[frame_plus_one].toks;
if (toks == NULL)
- KALDI_WARN << "No tokens alive [doing pruning]\n";
+ KALDI_WARN << "No tokens alive [doing pruning]";
Token *tok, *next_tok, *prev_tok = NULL;
for (tok = toks; tok != NULL; tok = next_tok) {
next_tok = tok->next;
diff --cc src/decoder/lattice-simple-decoder.cc
Simple merge
diff --cc src/decoder/lattice-tracking-decoder.cc
Simple merge
diff --cc src/feat/feature-functions.cc
Simple merge
diff --cc src/feat/pitch-functions-test.cc
index 3f93a94ec36617b0c615bcb35f974632784fee19,9c484b2b5012ec5611e516e36ed50d72ac641cec..642eff602c930e0fa9f40fd151c6b95c67a4e9cd
}
static void UnitTestSimple() {
- KALDI_LOG << "=== UnitTestSimple() ===\n";
+ KALDI_LOG << "=== UnitTestSimple() ===";
Vector<BaseFloat> v(1000);
- Vector<BaseFloat> out;
- Matrix<BaseFloat> m;
+ Matrix<BaseFloat> m1, m2;
// init with noise
for (int32 i = 0; i < v.Dim(); i++) {
v(i) = (abs(i * 433024253) % 65535) - (65535 / 2);
}
- KALDI_LOG << "<<<=== Just make sure it runs... Nothing is compared\n";
+ KALDI_LOG << "<<<=== Just make sure it runs... Nothing is compared";
- // the parametrization object
- PitchExtractionOptions op;
- // trying to have same opts as baseline.
- // compute pitch.
- Compute(op, v, &m);
+ // trying to compute and process pitch with same opts as baseline.
+ PitchExtractionOptions op1;
+ ProcessPitchOptions op2;
+ ComputeAndProcessKaldiPitch(op1, op2, v, &m1);
- KALDI_LOG << "Test passed :)\n";
+ KALDI_LOG << "Test passed :)";
}
+
+
+// Make sure that doing a calculation on the whole waveform gives
+// the same results as doing on the waveform broken into pieces.
+static void UnitTestPieces() {
+ KALDI_LOG << "=== UnitTestPieces() ===\n";
+ for (int32 n = 0; n < 10; n++) {
+ // the parametrization object
+ PitchExtractionOptions op1;
+ ProcessPitchOptions op2;
+ op2.delta_pitch_noise_stddev = 0.0; // to avoid mismatch of delta_log_pitch
+ // brought by rand noise.
+ op1.nccf_ballast_online = true; // this is necessary for the computation
+ // to be identical regardless how many pieces we break the signal into.
+
+ int32 size = 10000 + rand() % 50000;
+
+ Vector<BaseFloat> v(size);
+ // init with noise plus a sine-wave whose frequency is changing randomly.
+
+ double cur_freq = 200.0, normalized_time = 0.0;
+
+ for (int32 i = 0; i < size; i++) {
+ v(i) = RandGauss() + cos(normalized_time * M_2PI);
+ cur_freq += RandGauss(); // let the frequency wander a little.
+ if (cur_freq < 100.0) cur_freq = 100.0;
+ if (cur_freq > 300.0) cur_freq = 300.0;
+ normalized_time += cur_freq / op1.samp_freq;
+ }
+
+ Matrix<BaseFloat> m1, m1p;
+
+ // trying to have same opts as baseline.
+ ComputeKaldiPitch(op1, v, &m1);
+ ProcessPitch(op2, m1, &m1p);
+
+ Matrix<BaseFloat> m2, m2p;
+
+ { // compute it online with multiple pieces.
+ OnlinePitchFeature pitch_extractor(op1);
+ OnlineProcessPitch process_pitch(op2, &pitch_extractor);
+ int32 start_samp = 0;
+ while (start_samp < v.Dim()) {
+ int32 num_samp = rand() % (v.Dim() + 1 - start_samp);
+ SubVector<BaseFloat> v_part(v, start_samp, num_samp);
+ pitch_extractor.AcceptWaveform(op1.samp_freq, v_part);
+ start_samp += num_samp;
+ }
+ pitch_extractor.InputFinished();
+ int32 num_frames = pitch_extractor.NumFramesReady();
+ m2.Resize(num_frames, 2);
+ m2p.Resize(num_frames, process_pitch.Dim());
+ for (int32 frame = 0; frame < num_frames; frame++) {
+ SubVector<BaseFloat> row(m2, frame);
+ pitch_extractor.GetFrame(frame, &row);
+ SubVector<BaseFloat> rowp(m2p, frame);
+ process_pitch.GetFrame(frame, &rowp);
+ }
+ }
+ AssertEqual(m1, m2);
+ if (!ApproxEqual(m1p, m2p)) {
+ KALDI_ERR << "Post-processed pitch differs: " << m1p << " vs. " << m2p;
+ }
+ KALDI_LOG << "Test passed :)\n";
+ }
+}
+
+extern bool pitch_use_naive_search; // was declared in pitch-functions.cc
+
+// Make sure that doing a calculation on the whole waveform gives
+// the same results as doing on the waveform broken into pieces.
+static void UnitTestSearch() {
+ KALDI_LOG << "=== UnitTestSearch() ===\n";
+ for (int32 n = 0; n < 3; n++) {
+ // the parametrization object
+ PitchExtractionOptions op;
+ op.nccf_ballast_online = true; // this is necessary for the computation
+ // to be identical regardless how many pieces we break the signal into.
+
+ int32 size = 10000 + rand() % 10000;
+
+ Vector<BaseFloat> v(size);
+ // init with noise plus a sine-wave whose frequency is changing randomly.
+
+ double cur_freq = 200.0, normalized_time = 0.0;
+
+ for (int32 i = 0; i < size; i++) {
+ v(i) = RandGauss() + cos(normalized_time * M_2PI);
+ cur_freq += RandGauss(); // let the frequency wander a little.
+ if (cur_freq < 100.0) cur_freq = 100.0;
+ if (cur_freq > 300.0) cur_freq = 300.0;
+ normalized_time += cur_freq / op.samp_freq;
+ }
+
+ Matrix<BaseFloat> m1;
+ ComputeKaldiPitch(op, v, &m1);
+
+ pitch_use_naive_search = true;
+
+ Matrix<BaseFloat> m2;
+ ComputeKaldiPitch(op, v, &m2);
+
+ pitch_use_naive_search = false;
+
+ AssertEqual(m1, m2, 1.0e-08); // should be identical.
+ }
+ KALDI_LOG << "Test passed :)\n";
+}
+
+static void UnitTestComputeGPE() {
+ KALDI_LOG << "=== UnitTestComputeGPE ===\n";
+ int32 wrong_pitch = 0, tot_voiced = 0, tot_unvoiced = 0, num_frames = 0;
+ BaseFloat tol = 0.1, avg_d_kpitch = 0, real_pitch = 0;
+ for (int32 i = 1; i < 11; i++) {
+ std::string wavefile;
+ std::string num;
+ if (i < 6) {
+ num = "f" + ConvertIntToString(i) + "nw0000";
+ } else {
+ num = "m" + ConvertIntToString(i-5) + "nw0000";
+ }
+ Matrix<BaseFloat> gross_pitch;
+ std::string pitchfile = "keele/keele-true-lags/"+num+".txt";
+ std::ifstream pitch(pitchfile.c_str());
+ gross_pitch.Read(pitch, false);
+ Matrix<BaseFloat> kaldi_pitch;
+ std::string kfile = "keele/tmp/+"+num+"-kaldi.txt";
+ std::ifstream kpitch(kfile.c_str());
+ kaldi_pitch.Read(kpitch, false);
+ num_frames = std::min(kaldi_pitch.NumRows(),gross_pitch.NumRows());
+ for (int32 j = 1; j < num_frames; j++) {
+ if (gross_pitch(j,0) > 0.0) {
+ tot_voiced++;
+ real_pitch = 20000.0/gross_pitch(j,0);
+ if (fabs((real_pitch - kaldi_pitch(j,1))/real_pitch) > tol)
+ wrong_pitch++;
+ } else if (gross_pitch(j,0) == 0.0 && gross_pitch(j-1,0) == 0.0) {
+ tot_unvoiced++;
+ avg_d_kpitch += fabs(kaldi_pitch(j,1) - kaldi_pitch(j-1,1));
+ }
+ }
+ }
+ BaseFloat GPE = 1.0 * wrong_pitch / tot_voiced;
+ KALDI_LOG << " Gross Pitch Error with Rel.Error " << tol << " is " << GPE;
+ KALDI_LOG << "Average Kaldi delta_pitch for unvoiced regions " << avg_d_kpitch/tot_unvoiced;
+}
+
// Compare pitch using Kaldi pitch tracker on KEELE corpora
static void UnitTestKeele() {
- KALDI_LOG << "=== UnitTestKeele() ===\n";
+ KALDI_LOG << "=== UnitTestKeele() ===";
for (int32 i = 1; i < 11; i++) {
std::string wavefile;
std::string num;
}
}
}
-static void UnitTestWeightedMwn() {
- KALDI_LOG << "=== UnitTestWeightedMwn1() ===";
- // compare the results of WeightedMwn1 and Sliding CMN with uniform weights.
- for (int32 i = 0; i < 1000; i++) {
- int32 num_frames = 1 + (rand()%10 * 10);
- int32 normalization_win_size = 5 + rand() % 50;
- Matrix<BaseFloat> feat(num_frames, 2),
- output_feat(num_frames, 2);
- feat.SetRandn();
- for (int32 j = 0; j < num_frames; j++)
- feat(j, 0) = 1;
-
- Vector<BaseFloat> pov(num_frames),
- log_pitch(num_frames),
- mean_subtracted_log_pitch(num_frames);
- pov.CopyColFromMat(feat, 0);
- log_pitch.CopyColFromMat(feat, 1);
- WeightedMwn(normalization_win_size, pov, log_pitch ,
- &mean_subtracted_log_pitch);
- output_feat.CopyColFromVec(mean_subtracted_log_pitch, 1);
-
- // SlidingWindow
- SlidingWindowCmnOptions opts;
- opts.cmn_window = normalization_win_size;
- opts.center = true;
- opts.min_window = 1 + rand() % 100;
- if (opts.min_window > opts.cmn_window)
- opts.min_window = opts.cmn_window;
- Matrix<BaseFloat> output_feat2(num_frames, 2);
- SlidingWindowCmn(opts, feat, &output_feat2);
- for (int32 j = 0; j < num_frames; j++)
- output_feat(j, 0) = 0.0;
- if (!output_feat.ApproxEqual(output_feat2, 0.001)) {
- Matrix<BaseFloat> output_all(num_frames, 2);
- Vector<BaseFloat> pitch(num_frames),
- pitch2(num_frames);
- pitch.CopyColFromMat(output_feat, 1);
- pitch2.CopyColFromMat(output_feat2, 1);
- output_all.CopyColFromVec(pitch, 0);
- output_all.CopyColFromVec(pitch2, 1);
- KALDI_ERR << "Feafures differ:\n" << output_all;
- }
- }
- // Weighted Moving Window Normalization with non-uniform weights
- /*
- int32 num_frames = 1 + (rand()%10 * 20);
- int32 normalization_win_size = 5 + rand() % 50;
- Matrix<BaseFloat> feat(num_frames, 2),
- output_feat(num_frames, 2);
- for (int32 j = 0; j < num_frames; j++) {
- int32 r = rand() % 2;
- feat(j, 0) = RandUniform() / (1 + 1000.0 * r);
- feat(j, 1) = feat(j, 1) * feat(j, 0);
- }
- ProcessPovFeatures(&feat, 2, true);
- WeightedMwn(normalization_win_size, feat, &output_feat);
- */
-}
-static void UnitTestTakeLogOfPitch() {
- for (int32 i = 0; i < 100; i++) {
- int num_frame = 50 + (rand() % 200 * 200);
- Matrix<BaseFloat> input(num_frame, 2);
- input.SetRandn();
- input.Scale(100);
- Matrix<BaseFloat> output(input);
- for (int j = 0; j < num_frame; j++) {
- if (input(j, 1) < 1) {
- input(j, 1) = 10;
- output(j, 1) = 10;
- }
- output(j, 1) = log(input(j, 2));
- }
- TakeLogOfPitch(&input);
- if (input.ApproxEqual(output, 0.0001)) {
- KALDI_ERR << " Log of Matrix differs " << input << " vs. " << output;
- }
- }
-}
+
static void UnitTestPitchExtractionSpeed() {
- KALDI_LOG << "=== UnitTestPitchExtractionSpeed() ===\n";
+ KALDI_LOG << "=== UnitTestPitchExtractionSpeed() ===";
// use pitch code with default configuration..
PitchExtractionOptions op;
op.nccf_ballast = 0.1;
if (DirExist("keele/16kHz")) {
UnitTestFeatWithKeele();
} else {
- KALDI_LOG << "Not running tests that require the Keele database, "
- << "please ask g.meyer@somewhere.edu for the database if you need it.\n"
- << " you need to put keele wave file in keele/16kHz directory";
+ KALDI_LOG
+ << "Not running tests that require the Keele database, "
+ << "please ask g.meyer@liverpool.ac.uk for the database if you need it.\n"
+ << "Once you have the keele/ subdirectory, containing *.{pel,pet,pev,raw,wav}, do this:\n"
+ << "cd keele; mkdir -p 16kHz; mkdir -p tmp; for x in *.wav; do \n"
+ << "sox $x -r 16000 16kHz/$x; done \n"
+ << "mkdir -p keele-true-lags; for f in *.pev; do \n"
+ << "out_f=keele-true-lags/$(echo $f | sed s:pev:txt:); ( echo ' ['; len=`cat $f | wc -l`; \n"
+ << "head -n $(($len-1)) $f | tail -n $(($len-14)) ; echo -n ']') >$out_f; done \n"
+ << "\n"
+ << "Note: the GPE reported in paper is computed using pseudo-ground-truth pitch obtained\n"
+ << "by voting among the pitch trackers mentioned in the paper.\n";
}
- KALDI_LOG << "Tests succeeded.\n";
+ KALDI_LOG << "Tests succeeded.";
return 0;
} catch(const std::exception &e) {
KALDI_ERR << e.what();
diff --cc src/gmmbin/gmm-decode-faster.cc
Simple merge
diff --cc src/gmmbin/gmm-decode-nbest.cc
Simple merge
diff --cc src/hmm/posterior.cc
Simple merge
diff --cc src/ivector/ivector-extractor-test.cc
Simple merge
diff --cc src/lat/determinize-lattice-pruned.cc
Simple merge
diff --cc src/latbin/Makefile
index 7c7051b4011984d2467144a73bfb10ed63cc3418,300248cd367445c64127a682e3a455caaf5d4139..daa8e91fde24c66099843c5d6cae832e3ec15ab8
--- 1/src/latbin/Makefile
--- 2/src/latbin/Makefile
+++ b/src/latbin/Makefile
lattice-to-smbr-post lattice-determinize-pruned-parallel \
lattice-add-penalty lattice-align-words-lexicon lattice-push \
lattice-minimize lattice-limit-depth lattice-depth-per-frame \
- lattice-determinize-phone-pruned lattice-determinize-phone-pruned-parallel \
- lattice-expand-ngram
-
+ lattice-confidence lattice-determinize-phone-pruned \
- lattice-determinize-phone-pruned-parallel
-
++ lattice-determinize-phone-pruned-parallel lattice-expand-ngram
OBJFILES =
diff --cc src/matrix/optimization.cc
Simple merge
diff --cc src/nnet2/nnet-component.cc
Simple merge
diff --cc src/nnet2/nnet-component.h
Simple merge
diff --cc src/transform/cmvn.cc
Simple merge