Configurable low_freq high_freq, dithering (#664)

 function(download_kaldi_native_fbank)
   include(FetchContent)
 
-  set(kaldi_native_fbank_URL  "https://github.com/csukuangfj/kaldi-native-fbank/archive/refs/tags/v1.18.7.tar.gz")
-  set(kaldi_native_fbank_URL2 "https://huggingface.co/csukuangfj/sherpa-onnx-cmake-deps/resolve/main/kaldi-native-fbank-1.18.7.tar.gz")
-  set(kaldi_native_fbank_HASH "SHA256=e78fd9d481d83d7d6d1be0012752e6531cb614e030558a3491e3c033cb8e0e4e")
+  set(kaldi_native_fbank_URL  "https://github.com/csukuangfj/kaldi-native-fbank/archive/refs/tags/v1.19.1.tar.gz")
+  set(kaldi_native_fbank_URL2 "https://huggingface.co/csukuangfj/sherpa-onnx-cmake-deps/resolve/main/kaldi-native-fbank-1.19.1.tar.gz")
+  set(kaldi_native_fbank_HASH "SHA256=0cae8cbb9ea42916b214e088912f9e8f2f648f54756b305f93f552382f31f904")
 
   set(KALDI_NATIVE_FBANK_BUILD_TESTS OFF CACHE BOOL "" FORCE)
   set(KALDI_NATIVE_FBANK_BUILD_PYTHON OFF CACHE BOOL "" FORCE)

@@ -25,6 +25,19 @@ void FeatureExtractorConfig::Register(ParseOptions *po) {
 
   po->Register("feat-dim", &feature_dim,
                "Feature dimension. Must match the one expected by the model.");
+
+  po->Register("low-freq", &low_freq,
+               "Low cutoff frequency for mel bins");
+
+  po->Register("high-freq", &high_freq,
+               "High cutoff frequency for mel bins "
+               "(if <= 0, offset from Nyquist)");
+
+  po->Register("dither", &dither,
+               "Dithering constant (0.0 means no dither). "
+               "By default the audio samples are in range [-1,+1], "
+               "so 0.00003 is a good value, "
+               "equivalent to the default 1.0 from kaldi");
 }
 
 std::string FeatureExtractorConfig::ToString() const {
@@ -32,7 +45,10 @@ std::string FeatureExtractorConfig::ToString() const {
 
   os << "FeatureExtractorConfig(";
   os << "sampling_rate=" << sampling_rate << ", ";
-  os << "feature_dim=" << feature_dim << ")";
+  os << "feature_dim=" << feature_dim << ", ";
+  os << "low_freq=" << low_freq << ", ";
+  os << "high_freq=" << high_freq << ", ";
+  os << "dither=" << dither << ")";
 
   return os.str();
 }
@@ -40,7 +56,7 @@ std::string FeatureExtractorConfig::ToString() const {
 class FeatureExtractor::Impl {
  public:
   explicit Impl(const FeatureExtractorConfig &config) : config_(config) {
-    opts_.frame_opts.dither = 0;
+    opts_.frame_opts.dither = config.dither;
     opts_.frame_opts.snip_edges = config.snip_edges;
     opts_.frame_opts.samp_freq = config.sampling_rate;
     opts_.frame_opts.frame_shift_ms = config.frame_shift_ms;
@@ -50,13 +66,9 @@ class FeatureExtractor::Impl {
 
     opts_.mel_opts.num_bins = config.feature_dim;
 
-    // Please see
-    // https://github.com/lhotse-speech/lhotse/blob/master/lhotse/features/fbank.py#L27
-    // and
-    // https://github.com/k2-fsa/sherpa-onnx/issues/514
-    opts_.mel_opts.high_freq = -400;
+    opts_.mel_opts.high_freq = config.high_freq;
+    opts_.mel_opts.low_freq  = config.low_freq;
 
-    opts_.mel_opts.low_freq = config.low_freq;
     opts_.mel_opts.is_librosa = config.is_librosa;
 
     fbank_ = std::make_unique<knf::OnlineFbank>(opts_);

@@ -21,6 +21,27 @@ struct FeatureExtractorConfig {
   // Feature dimension
   int32_t feature_dim = 80;
 
+  // minimal frequency for Mel-filterbank, in Hz
+  float low_freq = 20.0f;
+
+  // maximal frequency of Mel-filterbank
+  // in Hz; negative value is subtracted from Nyquist freq.:
+  // i.e. for sampling_rate 16000 / 2 - 400 = 7600Hz
+  //
+  // Please see
+  // https://github.com/lhotse-speech/lhotse/blob/master/lhotse/features/fbank.py#L27
+  // and
+  // https://github.com/k2-fsa/sherpa-onnx/issues/514
+  float high_freq = -400.0f;
+
+  // dithering constant, useful for signals with hard-zeroes in non-speech parts
+  // this prevents large negative values in log-mel filterbanks
+  //
+  // In k2, audio samples are in range [-1..+1], in kaldi the range was
+  // [-32k..+32k], so the value 0.00003 is equivalent to kaldi default 1.0
+  //
+  float dither = 0.0f;  // dithering disabled by default
+
   // Set internally by some models, e.g., paraformer sets it to false.
   // This parameter is not exposed to users from the commandline
   // If true, the feature extractor expects inputs to be normalized to
@@ -31,7 +52,6 @@ struct FeatureExtractorConfig {
   bool snip_edges = false;
   float frame_shift_ms = 10.0f;   // in milliseconds.
   float frame_length_ms = 25.0f;  // in milliseconds.
-  int32_t low_freq = 20;
   bool is_librosa = false;
   bool remove_dc_offset = true;       // Subtract mean of wave before FFT.
   std::string window_type = "povey";  // e.g. Hamming window

@@ -72,6 +72,8 @@ class KeywordSpotterTransducerImpl : public KeywordSpotterImpl {
       unk_id_ = sym_["<unk>"];
     }
 
+    model_->SetFeatureDim(config.feat_config.feature_dim);
+
     InitKeywords();
 
     decoder_ = std::make_unique<TransducerKeywordDecoder>(
@@ -89,6 +91,8 @@ class KeywordSpotterTransducerImpl : public KeywordSpotterImpl {
       unk_id_ = sym_["<unk>"];
     }
 
+    model_->SetFeatureDim(config.feat_config.feature_dim);
+
     InitKeywords(mgr);
 
     decoder_ = std::make_unique<TransducerKeywordDecoder>(

@@ -90,6 +90,8 @@ class OnlineRecognizerTransducerImpl : public OnlineRecognizerImpl {
       unk_id_ = sym_["<unk>"];
     }
 
+    model_->SetFeatureDim(config.feat_config.feature_dim);
+
     if (config.decoding_method == "modified_beam_search") {
       if (!config_.hotwords_file.empty()) {
         InitHotwords();
@@ -123,6 +125,8 @@ class OnlineRecognizerTransducerImpl : public OnlineRecognizerImpl {
       unk_id_ = sym_["<unk>"];
     }
 
+    model_->SetFeatureDim(config.feat_config.feature_dim);
+
     if (config.decoding_method == "modified_beam_search") {
 #if 0
       // TODO(fangjun): Implement it

@@ -61,6 +61,16 @@ class OnlineTransducerModel {
    */
   virtual std::vector<Ort::Value> GetEncoderInitStates() = 0;
 
+  /** Set feature dim.
+   *
+   * This is used in `OnlineZipformer2TransducerModel`,
+   * to pass `feature_dim` for `GetEncoderInitStates()`.
+   *
+   * This has to be called before GetEncoderInitStates(), so the `encoder_embed`
+   * init state has the correct `embed_dim` of its output.
+   */
+  virtual void SetFeatureDim(int32_t feature_dim) { }
+
   /** Run the encoder.
    *
    * @param features  A tensor of shape (N, T, C). It is changed in-place.

@@ -403,7 +403,10 @@ OnlineZipformer2TransducerModel::GetEncoderInitStates() {
   }
 
   {
-    std::array<int64_t, 4> s{1, 128, 3, 19};
+    SHERPA_ONNX_CHECK_NE(feature_dim_, 0);
+    int32_t embed_dim = (((feature_dim_ - 1) / 2) - 1) / 2;
+    std::array<int64_t, 4> s{1, 128, 3, embed_dim};
+
     auto v = Ort::Value::CreateTensor<float>(allocator_, s.data(), s.size());
     Fill(&v, 0);
     ans.push_back(std::move(v));

@@ -37,6 +37,10 @@ class OnlineZipformer2TransducerModel : public OnlineTransducerModel {
 
   std::vector<Ort::Value> GetEncoderInitStates() override;
 
+  void SetFeatureDim(int32_t feature_dim) override {
+    feature_dim_ = feature_dim;
+  }
+
   std::pair<Ort::Value, std::vector<Ort::Value>> RunEncoder(
       Ort::Value features, std::vector<Ort::Value> states,
       Ort::Value processed_frames) override;
@@ -101,6 +105,7 @@ class OnlineZipformer2TransducerModel : public OnlineTransducerModel {
 
   int32_t context_size_ = 0;
   int32_t vocab_size_ = 0;
+  int32_t feature_dim_ = 0;
 };
 
 }  // namespace sherpa_onnx

@@ -11,10 +11,17 @@ namespace sherpa_onnx {
 static void PybindFeatureExtractorConfig(py::module *m) {
   using PyClass = FeatureExtractorConfig;
   py::class_<PyClass>(*m, "FeatureExtractorConfig")
-      .def(py::init<int32_t, int32_t>(), py::arg("sampling_rate") = 16000,
-           py::arg("feature_dim") = 80)
+      .def(py::init<int32_t, int32_t, float, float, float>(),
+           py::arg("sampling_rate") = 16000,
+           py::arg("feature_dim") = 80,
+           py::arg("low_freq") = 20.0f,
+           py::arg("high_freq") = -400.0f,
+           py::arg("dither") = 0.0f)
       .def_readwrite("sampling_rate", &PyClass::sampling_rate)
       .def_readwrite("feature_dim", &PyClass::feature_dim)
+      .def_readwrite("low_freq", &PyClass::low_freq)
+      .def_readwrite("high_freq", &PyClass::high_freq)
+      .def_readwrite("dither", &PyClass::high_freq)
       .def("__str__", &PyClass::ToString);
 }
 

@@ -41,6 +41,9 @@ class OnlineRecognizer(object):
         num_threads: int = 2,
         sample_rate: float = 16000,
         feature_dim: int = 80,
+        low_freq: float = 20.0,
+        high_freq: float = -400.0,
+        dither: float = 0.0,
         enable_endpoint_detection: bool = False,
         rule1_min_trailing_silence: float = 2.4,
         rule2_min_trailing_silence: float = 1.2,
@@ -80,6 +83,16 @@ class OnlineRecognizer(object):
             Sample rate of the training data used to train the model.
           feature_dim:
             Dimension of the feature used to train the model.
+          low_freq:
+            Low cutoff frequency for mel bins in feature extraction.
+          high_freq:
+            High cutoff frequency for mel bins in feature extraction
+            (if <= 0, offset from Nyquist)
+          dither:
+            Dithering constant (0.0 means no dither).
+            By default the audio samples are in range [-1,+1],
+            so dithering constant 0.00003 is a good value,
+            equivalent to the default 1.0 from kaldi
           enable_endpoint_detection:
             True to enable endpoint detection. False to disable endpoint
             detection.
@@ -140,6 +153,9 @@ class OnlineRecognizer(object):
         feat_config = FeatureExtractorConfig(
             sampling_rate=sample_rate,
             feature_dim=feature_dim,
+            low_freq=low_freq,
+            high_freq=high_freq,
+            dither=dither,
         )
 
         endpoint_config = EndpointConfig(