mvansegbroeck-zz / featxtra

Kaldi Speech Processing Tools

Featxtra

The featxtra toolbox lists a set of front-end tools and Signal Processing functions for Kaldi.

The toolbox includes:

• feature extraction for speech signals
  • Gammatone Frequency Representation, Gammatone Frequency Cepstral Coefficients
  • Gabor Features
  • DCT features (apply on ltsv, voicing stream)
• Voice Activity Detection
• additional signal filter operations

List of Functions

featbin

• featbin/compute-gtf-feats
  • extract Gammatone Frequency Representation (GTF)
  • extract Gammatone Frequency Cepstral Coefficients (GFCC)
• featbin/compute-gabor-feats
  • extract Gabor Features (GBF)
• featbin/compute-dct-feats
  • extract DCT Features from time domain signal
• featbin/apply-arma
  • apply Auto-Regressive Moving Average (ARMA) filtering on a spectral representation (e.g GTF )
• featbin/apply-ltsv
  • apply Long-Term Spectral Variabilityi (LTSV) stream on spectral representation (e.g GTF )
• featbin/apply-vad
  • apply Voice Activity Detection (VAD) using voicing and LTSV probability
• featbin/extract-dims
  • extract specified dimension range out of a feature matrix

feat

• feat/feature-gtf
  • GTF and GFCC feature implementation code

transform

• transform/featxtra-functions
  • additional signal processing functions

---

Instructions to run

1. Modify the Makefile in directories feat, featbin, and transform as follow:

• feat/Makefile should include feature-dctf.o feature-gtf.o feature-gabor.o under OBJFILES variable.
• featbin/Makefile should include compute-dctf-feats compute-gtf-feats compute-gabor-feats apply-arma apply-ltsv apply-nccf-to-pov apply-vad-merged apply-vad extract-dims under the BINFILES variable.
• transform/Makefile should include featxtra-functions.o under the OBJFILES variable.

2. Add the following code in the src/feat/feature-window.cc file

void ExtractWaveformRemainder(const VectorBase<BaseFloat> &wave,
                              const FrameExtractionOptions &opts,
                              Vector<BaseFloat> *wave_remainder) {
  int32 frame_shift = opts.WindowShift();
  int32 num_frames = NumFrames(wave.Dim(), opts);
  // offset is the amount at the start that has been extracted.
  int32 offset = num_frames * frame_shift;
  KALDI_ASSERT(wave_remainder != NULL);
  int32 remaining_len = wave.Dim() - offset;
  wave_remainder->Resize(remaining_len);
  KALDI_ASSERT(remaining_len >= 0);
  if (remaining_len > 0)
    wave_remainder->CopyFromVec(SubVector<BaseFloat>(wave, offset, remaining_len));
}

3. Add the following code in the src/featbin/feature-window.h file

// ExtractWaveformRemainder is useful if the waveform is coming in segments.
// It extracts the bit of the waveform at the end of this block that you
// would have to append the next bit of waveform to, if you wanted to have
// the same effect as everything being in one big block.
void ExtractWaveformRemainder(const VectorBase<BaseFloat> &wave,
                              const FrameExtractionOptions &opts,
                              Vector<BaseFloat> *wave_remainder);

Steps 2 and 3 are done because of this commit in the kaldi-asr project.

After these steps just run the make command again in the src directory to finally integrate these in your kaldi project.

featxtra's Issues

Unable to create .o files

Hi
Upon copying the files from feat, featbin, and transform directories to the respective directories in kaldi/src, and upon running the make command, it doesn't create a .o file for compute-gtf-feats.cc like there exists one for compute-mfcc-feats.cc which is compute-gtf-feats.o.

Kindly look into the issue.
Thanks

Invalid option (configuration)

Hi
When I use the featxtra/test/conf/gtf.conf file for compute-gabor-feats and compute-dctf-feats, it throws an invalid option error.

ERROR (compute-gabor-feats[5.5.103~4-22834]:ReadConfigFile():parse-options.cc:500) Invalid option --num-bins=64 in config file conf/dctf.conf

What should the .conf file contain to correct this error?

(Possibly) Incorrect values of GFCC feature matrices

Hi
When you run the command copy-matrix for converting features to matrix for mfcc, it shows 13 columns in that matrix corresponding to 13 features. Whereas that for GFCC gives a weird output (instead of showing 13 columns for 13 features). I have attached an image of that output.

Can you please verify whether the output is correct or not? And what should I do in case I want only 13 features extracted as in the case of MFCC?

How to integrate this with kaldi?

The instructions do not tell how to integrate this with existing set up. Please reply on how to replace the existing mfcc feature extraction technique with gfcc.

Thanks.

How to install?

Can you please let me know how to install it? Thank you very much.

"Could not read any features"

We want to replace mfcc with gfcc in kaldi . So When running Kaldi for dummies, on replacing compute-mfcc-feats with compute-gtf-feat in make_mfcc.sh file it gives the error in the monophone training phase.Kindly help in its resolution.Urgently needed!Thank you.