Signal processing for robust speech recognition

摘要: This paper describes a series of cepstral-based compensation procedures that render the SPHINX-II system more robust with respect to acoustical environment. The first algorithm, phone-dependent cepstral compensation, is similar in concept previously-described MFCDCN method, except vectors are selected according current phonetic hypothesis, rather than on basis SNR or VQ codeword identity. We also describe two accomplish adaptation codebook for new environments, as well use reduced-bandwidth frequency analysis process telephone-bandwidth speech. Use various algorithms consort produces reduction error rates by much 40 percent relative rate achieved mean normalization alone, both development test sets and context 1993 ARPA CSR evaluations.