Spectral

Python package for extracing Mel and MFCC features from speech.

• Free software: GPL3 license

Features

• Mel and MFCC feature extraction

Usage

To extract standard-ish MFCC features with deltas from a mono wave file:

>>> # read in the wave file
>>> import wave
>>> import struct
>>> with wave.open("mywavfile.wav", 'r') as fid:
        _, _, fs, nframes, _, _ = fid.getparams()
        sig = np.array(struct.unpack_from("%dh" % nframes,
                                          fid.readframes(nframes)))
>>> config = dict(fs=fs, scale='mel', do_dct=True, deltas=True)
>>> extractor = spectral.Spectral(**config)
>>> cepstra = extractor.transform(sig)

To extract 40-dimensional Bark filterbank features instead:

>>> config = dict(fs=fs, do_dct=False, scale='bark', deltas=False)
>>> extractor = spectral.Spectral(**config)
>>> bark_spectrum = extractor.transform(sig)

To extract the same features but first apply some noise removal:

>>> config = dict(fs=fs, do_dct=False, scale='bark', deltas=False,
                  remove_dc=True, medfilt_t=3, medfilt_s=(11,11), noise_fr=10)
>>> extractor = spectral.Spectral(**config)
>>> bark_spectrum = extractor.transform(sig)