Alain Biem, Shigeru Katagiri
A Study of Cepstrum Optimization by Discriminative Feature Extraction
- DFE implementation details -
Abstract:This report discusses application of Discriminative Feature Extraction (DFE) to speech recognition. The report is an opportunity to discuss DFE implementation on speech recognizers. The implementation process is viewed in detail in the context of filter-bank optimization. Choice of learning rate parameters, smoothing of the loss, optimization methodology are described for an accurate application to speech recognition. As illustration, application of DFE to cepstrum optimization is studied, for a multi-speaker vowel recognition task. It is shown that DFE-based cepstrum are more robust than conventional Mel-scale based cepstrum coefficients.