Normfeat

Copyright 2014-2020 Anthony Larcher and Sylvain Meignier

frontend provides methods to process an audio signal in order to extract useful parameters for speaker verification.

frontend.normfeat.cep_sliding_norm(features, win=301, label=None, center=True, reduce=False)[source]

Performs a cepstal mean substitution and standard deviation normalization in a sliding windows. MFCC is modified.

Parameters

  • features – the MFCC, a numpy array

  • win – the size of the sliding windows

  • label – vad label if available

  • center – performs mean subtraction

  • reduce – performs standard deviation division

frontend.normfeat.cms(features, label=None, global_mean=None)[source]

Performs cepstral mean subtraction

Parameters

  • features – a feature stream of dimension dim x nframes where dim is the dimension of the acoustic features and nframes the number of frames in the stream

  • label – a logical vector

  • global_mean – pre-computed mean to use for feature normalization if given

Returns

a feature stream

frontend.normfeat.cmvn(features, label=None, global_mean=None, global_std=None)[source]

Performs mean and variance normalization

Parameters

  • features – a feature stream of dimension dim x nframes where dim is the dimension of the acoustic features and nframes the number of frames in the stream

  • global_mean – pre-computed mean to use for feature normalization if given

  • global_std – pre-computed standard deviation to use for feature normalization if given

  • label – a logical verctor

Returns

a sequence of features

frontend.normfeat.rasta_filt(x)[source]

Apply RASTA filtering to the input signal.

Parameters

x – the input audio signal to filter. cols of x = critical bands, rows of x = frame same for y but after filtering default filter is single pole at 0.94

frontend.normfeat.stg(features, label=None, win=301)[source]

Performs feature warping on a sliding window

Parameters

  • features – a feature stream of dimension dim x nframes where dim is the dimension of the acoustic features and nframes the number of frames in the stream

  • label – label of selected frames to compute the Short Term Gaussianization, by default, al frames are used

  • win – size of the frame window to consider, must be an odd number to get a symetric context on left and right

Returns

a sequence of features