change documentation of --dither option

cmake/kaldi-native-fbank.cmake

@@ -1,6 +1,7 @@
 function(download_kaldi_native_fbank)
   include(FetchContent)
 
+  # TODO: update is required, so that dithering works... (it was missing)
   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")

sherpa-onnx/csrc/features.cc

@@ -36,7 +36,8 @@ void FeatureExtractorConfig::Register(ParseOptions *po) {
   po->Register("dither", &dither,
                "Dithering constant (0.0 means no dither). "
                "By default the audio samples are in range [-1,+1], "
-               "so dithering constant should be around 0.0001");
+               "so 0.00003 is a good value, "
+               "equivalent to the default 1.0 from kaldi");
 }
 
 std::string FeatureExtractorConfig::ToString() const {

sherpa-onnx/csrc/features.h

@@ -37,10 +37,10 @@ struct FeatureExtractorConfig {
   // dithering constant, useful for signals with hard-zeroes in non-speech parts
   // this prevents large negative values in log-mel filterbanks
   //
-  // By default the audio samples are in range [-1,+1],
-  // so dithering constant should be around 0.0001
+  // 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;
+  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