Approaches to Microphone Independence in Automatic Speech Recognition

This paper describes a series of cepstral-based compensation procedures that render the SPHINX-II system more robust with respect to acoustical changes in the environment. The first algorithm, RATZ (MultivaRiate gAussian based cepsTral normaliZation) requires stereo-data for computing compensation terms, and is similar in philosophy to MFCDCN [ref] (in fact MFCDCN can be thought of as a discrete case of RATZ). We also describe a second algorithm, an improved version of CDCN, that does not require stereo training data and yet achieves performance levels comparable to the RATZ and other stereo algorithms. Use of the various compensation algorithms in consort produces a reduction of error rates for SPHINX-II by as much as 20.0% percent relative to the rate achieved with cepstral mean normalization alone, in both development test sets and in the context of the 1994 ARPA CSR evaluations.