Missing-data methods with cepstral data

Andrew Morris

IDIAP, Martigny, Switzerland

Noise robust ASR is achieved by the MD ASR approach, as developed at Sheffield, through the exploitation of spectral redundancy, together with techniques for detecting clean spectral data. While models based on data which has been orthogonalised (via DCT or PCA) are more accurate, MD has until now been confined to models based on spectral data. So the question arrises: can the advantages of spectral redundancy and orthogonalised data be combined? A number of ideas relating to this have been tested recently at IDIAP.

1. use localised MLPs to generate "data utility masks" for any data representation. So far these have not led to improved recognition performance.
2. propagate spectral intervals of uncertainty through the DCT to obtain corresponding cepstral intervals. Result is hugh intervals on all cepstral coeffs, except when no spectral data is missing in one frame.
3. introduce redundancy into cepstral domain by appending features obtained through separate application of DCT to top and bottom half of frequency range. Still very large intervals, and no improvement over cepstral baseline.
4. restrict cepstral intervals by restricting spectral intervals through refinements relative to the "maximum assumption". Can greataly reduce cepstral uncertainty, but becomes complex, and so far no improvement in cepstral performance.