- kaldifeats.commands:
A wrapper in python modules to invoke commad line tools of Kaldi
, ensuring 100% reproducibility of kaldi features, of course.
- kaldifeats.signal:
Pure python implementations to perform stft
.
Actually, instead of Kaldi, it depends on scipy for some useful signal funciton, e.g. scipy.signal.get_window
,
but not using the highest function, scipy.signal.stft
, scipy.signal.spectrogram
.
- kaldifeats.feats:
Pure python modules for audio feature extractor, e.g. log-fbank-feats
, mfcc
, add-deltas
.
python3.6
scipy
numpy
kaldiio (pip install git+https://github.com/nttcslab-sp/kaldiio)
- https://github.com/scipy/scipy/blob/master/scipy/signal/spectral.py
- https://github.com/kaldi-asr/kaldi/blob/master/src/feat/feature-window.cc
- https://github.com/kaldi-asr/kaldi/blob/master/src/feat/feature-spectrogram.cc
-
https://github.com/librosa/librosa
- Depends on scipy and audioread
-
https://github.com/MTG/essentia
- C++ implementation and python binding
- Supports loader for audio file depending libavcodec/libavformat/libavutil/libavresample(FFmpeg/LibAv)
- Not supporting Python3 (MTG/essentia#138)
-
https://github.com/jameslyons/python_speech_features
- Simple implementation