Software

• GitHub link

• All-pole group delay features (alternative to MFCCs for feature extraction from speech signals)

  • See the 2013 Interspeech paper and the Matlab code

  • Usage:

    featureVectors = extract_lpGdelayVec_diff(speechFrames, LP_ORDER, NFFT);

    • The featureVectors matrix has one row per frame of speech. featureVectors is a MxNFFT matrix.

    • speechFrames is a MxN matrix with M frames, each of N samples.

    • To obtain "usual" features, take a DCT of size P on the featureVectors matrix. P can be, for instance 18, resulting in an 18-dimensional feature vector per frame.

    • LP_ORDER is the order of the all-pole model, we have usually used 20 for NIST speaker id data.

    • NFFT is FFT order.