rename interface & add comment to Dim()

speechx/examples/feat/linear_spectrogram_main.cc

@@ -25,10 +25,9 @@
 #include "kaldi/util/kaldi-io.h"
 #include "kaldi/util/table-types.h"
 
-DEFINE_string(wav_rspecifier, "", "test wav path");
-DEFINE_string(feature_wspecifier, "", "test wav ark");
-DEFINE_string(feature_check_wspecifier, "", "test wav ark");
-DEFINE_string(cmvn_write_path, "./cmvn.ark", "test wav ark");
+DEFINE_string(wav_rspecifier, "", "test wav scp path");
+DEFINE_string(feature_wspecifier, "", "output feats wspecifier");
+DEFINE_string(cmvn_write_path, "./cmvn.ark", "write cmvn");
 
 
 std::vector<float> mean_{
@@ -165,10 +164,10 @@ int main(int argc, char* argv[]) {
     // test feature linear_spectorgram: wave --> decibel_normalizer --> hanning
     // window -->linear_spectrogram --> cmvn
     int32 num_done = 0, num_err = 0;
-    // std::unique_ptr<ppspeech::FeatureExtractorInterface> data_source(new
-    // ppspeech::RawDataSource());
+    //std::unique_ptr<ppspeech::FeatureExtractorInterface> data_source(new
+     //ppspeech::RawDataCache());
     std::unique_ptr<ppspeech::FeatureExtractorInterface> data_source(
-        new ppspeech::RawAudioSource());
+        new ppspeech::RawAudioCache());
 
     ppspeech::LinearSpectrogramOptions opt;
     opt.frame_opts.frame_length_ms = 20;
@@ -211,7 +210,7 @@ int main(int argc, char* argv[]) {
                 wav_chunk(i) = waveform(sample_offset + i);
             }
             kaldi::Vector<BaseFloat> features;
-            feature_cache.AcceptWaveform(wav_chunk);
+            feature_cache.Accept(wav_chunk);
             if (cur_chunk_size < chunk_sample_size) {
                 feature_cache.SetFinished();
             }

speechx/speechx/frontend/feature_cache.cc

@@ -29,9 +29,9 @@ FeatureCache::FeatureCache(
     base_extractor_ = std::move(base_extractor);
 }
 
-void FeatureCache::AcceptWaveform(
-    const kaldi::VectorBase<kaldi::BaseFloat>& input) {
-    base_extractor_->AcceptWaveform(input);
+void FeatureCache::Accept(
+    const kaldi::VectorBase<kaldi::BaseFloat>& inputs) {
+    base_extractor_->Accept(inputs);
     // feed current data
     bool result = false;
     do {
@@ -40,7 +40,7 @@ void FeatureCache::AcceptWaveform(
 }
 
 // pop feature chunk
-bool FeatureCache::Read(kaldi::Vector<kaldi::BaseFloat>* feat) {
+bool FeatureCache::Read(kaldi::Vector<kaldi::BaseFloat>* output_feats) {
     kaldi::Timer timer;
     std::unique_lock<std::mutex> lock(mutex_);
     while (cache_.empty() && base_extractor_->IsFinished() == false) {
@@ -53,8 +53,8 @@ bool FeatureCache::Read(kaldi::Vector<kaldi::BaseFloat>* feat) {
         usleep(1000);  // sleep 1 ms
     }
     if (cache_.empty()) return false;
-    feat->Resize(cache_.front().Dim());
-    feat->CopyFromVec(cache_.front());
+    output_feats->Resize(cache_.front().Dim());
+    output_feats->CopyFromVec(cache_.front());
     cache_.pop();
     ready_feed_condition_.notify_one();
     return true;

speechx/speechx/frontend/feature_cache.h

@@ -24,12 +24,15 @@ class FeatureCache : public FeatureExtractorInterface {
     explicit FeatureCache(
         int32 max_size = kint16max,
         std::unique_ptr<FeatureExtractorInterface> base_extractor = NULL);
-    virtual void AcceptWaveform(
-        const kaldi::VectorBase<kaldi::BaseFloat>& input);
-    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* feat);
+    virtual void Accept(
+        const kaldi::VectorBase<kaldi::BaseFloat>& inputs);
+    // output_feats dim = num_frames * feature_dim
+    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* output_feats);
+    // feature cache only cache feature which from base extractor
     virtual size_t Dim() const { return base_extractor_->Dim(); }
     virtual void SetFinished() {
         base_extractor_->SetFinished();
+        // read the last chunk data
         Compute();
     }
     virtual bool IsFinished() const { return base_extractor_->IsFinished(); }
@@ -44,7 +47,7 @@ class FeatureCache : public FeatureExtractorInterface {
     std::unique_ptr<FeatureExtractorInterface> base_extractor_;
     std::condition_variable ready_feed_condition_;
     std::condition_variable ready_read_condition_;
-    //    DISALLOW_COPY_AND_ASSGIN(FeatureCache);
+    //DISALLOW_COPY_AND_ASSGIN(FeatureCache);
 };
 
 }  // namespace ppspeech

speechx/speechx/frontend/feature_extractor_interface.h

@@ -21,13 +21,18 @@ namespace ppspeech {
 
 class FeatureExtractorInterface {
   public:
-    virtual void AcceptWaveform(
-        const kaldi::VectorBase<kaldi::BaseFloat>& input) = 0;
-    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* feat) = 0;
+    // accept input data
+    virtual void Accept(
+        const kaldi::VectorBase<kaldi::BaseFloat>& inputs) = 0;
+    // get the processed result
+    // the length of output = feature_row * feature_dim,
+    // the Matrix is squashed into Vector
+    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* outputs) = 0;
+    // the Dim is the feature dim
     virtual size_t Dim() const = 0;
     virtual void SetFinished() = 0;
     virtual bool IsFinished() const = 0;
     // virtual void Reset();
 };
 
-}  // namespace ppspeech
+}  // namespace ppspeech
\ No newline at end of file

speechx/speechx/frontend/linear_spectrogram.cc

@@ -66,11 +66,11 @@ LinearSpectrogram::LinearSpectrogram(
     dim_ = fft_points_ / 2 + 1;  // the dimension is Fs/2 Hz
 }
 
-void LinearSpectrogram::AcceptWaveform(const VectorBase<BaseFloat>& input) {
-    base_extractor_->AcceptWaveform(input);
+void LinearSpectrogram::Accept(const VectorBase<BaseFloat>& inputs) {
+    base_extractor_->Accept(inputs);
 }
 
-bool LinearSpectrogram::Read(Vector<BaseFloat>* feat) {
+bool LinearSpectrogram::Read(Vector<BaseFloat>* output_feats) {
     Vector<BaseFloat> input_feats(chunk_sample_size_);
     bool flag = base_extractor_->Read(&input_feats);
     if (flag == false || input_feats.Dim() == 0) return false;
@@ -83,9 +83,10 @@ bool LinearSpectrogram::Read(Vector<BaseFloat>* feat) {
     if (result.size() != 0) {
         feat_size = result.size() * result[0].size();
     }
-    feat->Resize(feat_size);
+    output_feats->Resize(feat_size);
+    // todo refactor (SimleGoat)
     for (size_t idx = 0; idx < feat_size; ++idx) {
-        (*feat)(idx) = result[idx / dim_][idx % dim_];
+        (*output_feats)(idx) = result[idx / dim_][idx % dim_];
     }
     return true;
 }
@@ -117,7 +118,7 @@ bool LinearSpectrogram::NumpyFft(vector<BaseFloat>* v,
     return true;
 }
 
-// Compute spectrogram feat, only for test, remove later
+// Compute spectrogram feat
 // todo: refactor later (SmileGoat)
 bool LinearSpectrogram::Compute(const vector<float>& wave,
                                 vector<vector<float>>& feat) {

speechx/speechx/frontend/linear_spectrogram.h

@@ -38,9 +38,10 @@ class LinearSpectrogram : public FeatureExtractorInterface {
     explicit LinearSpectrogram(
         const LinearSpectrogramOptions& opts,
         std::unique_ptr<FeatureExtractorInterface> base_extractor);
-    virtual void AcceptWaveform(
-        const kaldi::VectorBase<kaldi::BaseFloat>& input);
-    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* feat);
+    virtual void Accept(
+        const kaldi::VectorBase<kaldi::BaseFloat>& inputs);
+    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* output_feats);
+    // the dim_ is the dim of single frame feature
     virtual size_t Dim() const { return dim_; }
     virtual void SetFinished() { base_extractor_->SetFinished(); }
     virtual bool IsFinished() const { return base_extractor_->IsFinished(); }
@@ -49,8 +50,6 @@ class LinearSpectrogram : public FeatureExtractorInterface {
     void Hanning(std::vector<kaldi::BaseFloat>* data) const;
     bool Compute(const std::vector<kaldi::BaseFloat>& wave,
                  std::vector<std::vector<kaldi::BaseFloat>>& feat);
-    void Compute(const kaldi::VectorBase<kaldi::BaseFloat>& input,
-                 kaldi::VectorBase<kaldi::BaseFloat>* feature);
     bool NumpyFft(std::vector<kaldi::BaseFloat>* v,
                   std::vector<kaldi::BaseFloat>* real,
                   std::vector<kaldi::BaseFloat>* img) const;
@@ -60,7 +59,6 @@ class LinearSpectrogram : public FeatureExtractorInterface {
     std::vector<kaldi::BaseFloat> hanning_window_;
     kaldi::BaseFloat hanning_window_energy_;
     LinearSpectrogramOptions opts_;
-    kaldi::Vector<kaldi::BaseFloat> waveform_;  // remove later, todo(SmileGoat)
     std::unique_ptr<FeatureExtractorInterface> base_extractor_;
     int chunk_sample_size_;
     DISALLOW_COPY_AND_ASSIGN(LinearSpectrogram);

speechx/speechx/frontend/normalizer.cc

@@ -31,23 +31,20 @@ DecibelNormalizer::DecibelNormalizer(
     std::unique_ptr<FeatureExtractorInterface> base_extractor) {
     base_extractor_ = std::move(base_extractor);
     opts_ = opts;
-    dim_ = 0;
+    dim_ = 1;
 }
 
-void DecibelNormalizer::AcceptWaveform(
-    const kaldi::VectorBase<BaseFloat>& input) {
-    // dim_ = input.Dim();
-    // waveform_.Resize(input.Dim());
-    // waveform_.CopyFromVec(input);
-    base_extractor_->AcceptWaveform(input);
+void DecibelNormalizer::Accept(
+    const kaldi::VectorBase<BaseFloat>& inputs_wave) {
+    base_extractor_->Accept(inputs_wave);
 }
 
-bool DecibelNormalizer::Read(kaldi::Vector<BaseFloat>* feat) {
-    // if (waveform_.Dim() == 0) return;
-    if (base_extractor_->Read(feat) == false || feat->Dim() == 0) {
+bool DecibelNormalizer::Read(kaldi::Vector<BaseFloat>* outputs_wave) {
+    if (base_extractor_->Read(outputs_wave) == false || 
+        outputs_wave->Dim() == 0) {
         return false;
     }
-    Compute(feat);
+    Compute(outputs_wave);
     return true;
 }
 
@@ -70,7 +67,7 @@ void CopyStdVector2Vector(const vector<BaseFloat>& input,
     }
 }
 
-bool DecibelNormalizer::Compute(VectorBase<BaseFloat>* feat) const {
+bool DecibelNormalizer::Compute(VectorBase<BaseFloat>* feats) const {
     // calculate db rms
     BaseFloat rms_db = 0.0;
     BaseFloat mean_square = 0.0;
@@ -78,9 +75,9 @@ bool DecibelNormalizer::Compute(VectorBase<BaseFloat>* feat) const {
     BaseFloat wave_float_normlization = 1.0f / (std::pow(2, 16 - 1));
 
     vector<BaseFloat> samples;
-    samples.resize(feat->Dim());
+    samples.resize(feats->Dim());
     for (size_t i = 0; i < samples.size(); ++i) {
-        samples[i] = (*feat)(i);
+        samples[i] = (*feats)(i);
     }
 
     // square
@@ -110,7 +107,7 @@ bool DecibelNormalizer::Compute(VectorBase<BaseFloat>* feat) const {
         item *= std::pow(10.0, gain / 20.0);
     }
 
-    CopyStdVector2Vector(samples, feat);
+    CopyStdVector2Vector(samples, feats);
     return true;
 }
 
@@ -124,16 +121,16 @@ CMVN::CMVN(std::string cmvn_file,
     dim_ = stats_.NumCols() - 1;
 }
 
-void CMVN::AcceptWaveform(const kaldi::VectorBase<kaldi::BaseFloat>& input) {
-    base_extractor_->AcceptWaveform(input);
+void CMVN::Accept(const kaldi::VectorBase<kaldi::BaseFloat>& feats) {
+    base_extractor_->Accept(feats);
     return;
 }
 
-bool CMVN::Read(kaldi::Vector<BaseFloat>* feat) {
-    if (base_extractor_->Read(feat) == false) {
+bool CMVN::Read(kaldi::Vector<BaseFloat>* outputs) {
+    if (base_extractor_->Read(outputs) == false) {
         return false;
     }
-    Compute(feat);
+    Compute(outputs);
     return true;
 }
 

speechx/speechx/frontend/normalizer.h

@@ -45,15 +45,16 @@ class DecibelNormalizer : public FeatureExtractorInterface {
     explicit DecibelNormalizer(
         const DecibelNormalizerOptions& opts,
         std::unique_ptr<FeatureExtractorInterface> base_extractor);
-    virtual void AcceptWaveform(
-        const kaldi::VectorBase<kaldi::BaseFloat>& input);
-    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* feat);
+    virtual void Accept(
+        const kaldi::VectorBase<kaldi::BaseFloat>& inputs_wave);
+    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* outputs_wave);
+    // noramlize audio, the dim is 1.
     virtual size_t Dim() const { return dim_; }
     virtual void SetFinished() { base_extractor_->SetFinished(); }
     virtual bool IsFinished() const { return base_extractor_->IsFinished(); }
 
   private:
-    bool Compute(kaldi::VectorBase<kaldi::BaseFloat>* feat) const;
+    bool Compute(kaldi::VectorBase<kaldi::BaseFloat>* feats) const;
     DecibelNormalizerOptions opts_;
     size_t dim_;
     std::unique_ptr<FeatureExtractorInterface> base_extractor_;
@@ -65,15 +66,19 @@ class CMVN : public FeatureExtractorInterface {
   public:
     explicit CMVN(std::string cmvn_file,
                   std::unique_ptr<FeatureExtractorInterface> base_extractor);
-    virtual void AcceptWaveform(
-        const kaldi::VectorBase<kaldi::BaseFloat>& input);
-    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* feat);
+    virtual void Accept(
+        const kaldi::VectorBase<kaldi::BaseFloat>& feats);
+
+    // the length of outputs = feature_row * feature_dim,
+    // the Matrix is squashed into Vector
+    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* outputs);
+    // the dim_ is the feautre dim.
     virtual size_t Dim() const { return dim_; }
     virtual void SetFinished() { base_extractor_->SetFinished(); }
     virtual bool IsFinished() const { return base_extractor_->IsFinished(); }
 
   private:
-    void Compute(kaldi::VectorBase<kaldi::BaseFloat>* feat) const;
+    void Compute(kaldi::VectorBase<kaldi::BaseFloat>* feats) const;
     void ApplyCMVN(kaldi::MatrixBase<BaseFloat>* feats);
     kaldi::Matrix<double> stats_;
     std::unique_ptr<FeatureExtractorInterface> base_extractor_;

speechx/speechx/frontend/raw_audio.cc

@@ -21,33 +21,25 @@ using kaldi::BaseFloat;
 using kaldi::VectorBase;
 using kaldi::Vector;
 
-RawAudioSource::RawAudioSource(int buffer_size)
+RawAudioCache::RawAudioCache(int buffer_size)
     : finished_(false), data_length_(0), start_(0), timeout_(1) {
     ring_buffer_.resize(buffer_size);
 }
 
-void RawAudioSource::AcceptWaveform(const VectorBase<BaseFloat>& data) {
+void RawAudioCache::Accept(const VectorBase<BaseFloat>& input_audio) {
     std::unique_lock<std::mutex> lock(mutex_);
-    while (data_length_ + data.Dim() > ring_buffer_.size()) {
+    while (data_length_ + input_audio.Dim() > ring_buffer_.size()) {
         ready_feed_condition_.wait(lock);
     }
-    for (size_t idx = 0; idx < data.Dim(); ++idx) {
-        ring_buffer_[idx % ring_buffer_.size()] = data(idx);
+    for (size_t idx = 0; idx < input_audio.Dim(); ++idx) {
+        int32 buffer_idx = (idx + start_) % ring_buffer_.size(); 
+        ring_buffer_[buffer_idx] = input_audio(idx);
     }
-    data_length_ += data.Dim();
+    data_length_ += input_audio.Dim();
 }
 
-// bool RawAudioSource::AcceptWaveform(BaseFloat* data, int length) {
-// std::unique_lock<std::mutex> lock(mutex_);
-// for (size_t idx = 0; idx < length; ++idx) {
-// ring_buffer_[idx % ring_buffer_.size()] = data[idx];
-//}
-// data_length_ += length;
-// finish_condition_.notify_one();
-//}
-
-bool RawAudioSource::Read(Vector<BaseFloat>* feat) {
-    size_t chunk_size = feat->Dim();
+bool RawAudioCache::Read(Vector<BaseFloat>* output_audio) {
+    size_t chunk_size = output_audio->Dim();
     kaldi::Timer timer;
     std::unique_lock<std::mutex> lock(mutex_);
     while (chunk_size > data_length_) {
@@ -69,11 +61,12 @@ bool RawAudioSource::Read(Vector<BaseFloat>* feat) {
     // read last chunk data
     if (chunk_size > data_length_) {
         chunk_size = data_length_;
-        feat->Resize(chunk_size);
+        output_audio->Resize(chunk_size);
     }
 
     for (size_t idx = 0; idx < chunk_size; ++idx) {
-        feat->Data()[idx] = ring_buffer_[idx];
+        int buff_idx = (start_ + idx) % ring_buffer_.size();
+        output_audio->Data()[idx] = ring_buffer_[buff_idx];
     }
     data_length_ -= chunk_size;
     start_ = (start_ + chunk_size) % ring_buffer_.size();

speechx/speechx/frontend/raw_audio.h

@@ -20,12 +20,13 @@
 
 namespace ppspeech {
 
-class RawAudioSource : public FeatureExtractorInterface {
+class RawAudioCache : public FeatureExtractorInterface {
   public:
-    explicit RawAudioSource(int buffer_size = kint16max);
-    virtual void AcceptWaveform(const kaldi::VectorBase<BaseFloat>& data);
-    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* feat);
-    virtual size_t Dim() const { return data_length_; }
+    explicit RawAudioCache(int buffer_size = kint16max);
+    virtual void Accept(const kaldi::VectorBase<BaseFloat>& input_audio);
+    virtual bool Read(kaldi::Vector<kaldi::BaseFloat>* output_audio);
+    // the audio dim is 1
+    virtual size_t Dim() const { return 1; }
     virtual void SetFinished() {
         std::lock_guard<std::mutex> lock(mutex_);
         finished_ = true;
@@ -41,14 +42,14 @@ class RawAudioSource : public FeatureExtractorInterface {
     std::condition_variable ready_feed_condition_;
     kaldi::int32 timeout_;
 
-    DISALLOW_COPY_AND_ASSIGN(RawAudioSource);
+    DISALLOW_COPY_AND_ASSIGN(RawAudioCache);
 };
 
 // it is a datasource for testing different frontend module.
-class RawDataSource : public FeatureExtractorInterface {
+class RawDataCache: public FeatureExtractorInterface {
   public:
-    explicit RawDataSource() { finished_ = false; }
-    virtual void AcceptWaveform(
+    explicit RawDataCache() { finished_ = false; }
+    virtual void Accept(
         const kaldi::VectorBase<kaldi::BaseFloat>& input) {
         data_ = input;
     }
@@ -60,6 +61,7 @@ class RawDataSource : public FeatureExtractorInterface {
         data_.Resize(0);
         return true;
     }
+    //the dim is data_ length
     virtual size_t Dim() const { return data_.Dim(); }
     virtual void SetFinished() { finished_ = true; }
     virtual bool IsFinished() const { return finished_; }
@@ -68,7 +70,7 @@ class RawDataSource : public FeatureExtractorInterface {
     kaldi::Vector<kaldi::BaseFloat> data_;
     bool finished_;
 
-    DISALLOW_COPY_AND_ASSIGN(RawDataSource);
+    DISALLOW_COPY_AND_ASSIGN(RawDataCache);
 };
 
 }  // namespace ppspeech