MVDR based feature extraction for speech recognition

A method for extracting feature vectors with reduced variance from a digitized utterance, comprising the step of:

• computing the spectral envelope estimates for each of at least two overlapping segments corresponding to a current overlapping frame in a digitized utterance,
• wherein said computing step comprises the steps of:
• averaging the spectral envelope estimates for the at least two overlapping segments to obtain a smoothed spectral envelope for the current overlapping frame;
• and
• computing a Mel-Filtered Cepstral Coefficients (MFCC) vector from the smoothed spectral envelope.

A method for extracting feature vectors with reduced variance from a digitized utterance, comprising the step of:

• computing the spectral envelope estimate from each of at least two overlapping segments corresponding to a current overlapping frame of a digitized utterance,
• wherein said computing step comprises the steps of:
• computing a Mel-Filtered Cepstral Coefficients (MFCC) vector for each of the at least two segments corresponding to the current overlapping frame, from the spectral envelope estimate;
• and
• averaging together the MFCC vector for each of the at least two overlapping segments to obtain a smoothed MFCC vector for the current overlapping frame.

A method for extracting feature vectors from a digitized utterance, comprising the steps of:

• segmenting the digitized utterance into overlapping frames that comprise data samples;
• for each of the overlapping frames,
• splitting the data samples from a start of a current overlapping frame to a start of a next overlapping frame into at least two overlapping segments;
• for each of the at least two overlapping segments,
• computing a Fast Fourier Transform (FFT) of Minimum Variance Distortionless Response (MVDR) coefficients corresponding to the at least two overlapping segments to generate a sampled version of a discrete-time Fourier Transform of the MVDR coefficients;
• placing the sampled version of the discrete-time Fourier Transform into one of a plurality of overlapping bins;
• for each of the plurality of overlapping bins,
• determining a weighted sum of an inverse of a magnitude of sampled versions of the discrete-time Fourier Transform placed therein to obtain a bin value therefore;
• calculating a Discrete Cosine Transform (DCT) of logarithms of bin values in the plurality of bins to produce a feature vector;
• and
• averaging feature vectors obtained from each of the at least two segments to obtain an averaged feature vector for the current overlapping frame.