A VQ‐based speaker normalization method and its applications to speech recognition

Abstract

A speaker‐specific vector quantization (VQ) codebook was proposed and applied successfully to both text‐independent and text‐dependent speaker recognition applications [e.g., F. K. Soong, A. E. Rosenberg, L. R. Rabiner, and B‐H. Juang, ICASSP‐85 (1985) and A. E. Rosenberg and F. K. Soong, ICASSP‐86 (1986)]. In this talk the same VQ codebook is used for constructing a transformation, or more precisely, a mapping, between the feature space of a new speaker and that of a standard speaker. The mapping is constructed by using standard dynamic programming (DP) procedure to align training tokens spoken by a new speaker with tokens of the same text spoken by the standard speaker. Along the optimal alignment paths, a correspondence between the spectral feature vectors of the new speaker and the VQ codebook indices of the standard speaker is established and, for each VQ codebook index, a centroid is computed as the “average” of all the corresponding feature vectors of the new speaker. A new VQ codebook for the new speaker is therefore generated, and it is a one‐to‐one projected image of the VQ codebook of the standard speaker onto the feature space of the new speaker. This speaker normalization procedure has various applications in speech signal processing such as speech recognition, speech coding, and text‐to‐speech synthesis, etc. In this talk, isolated word recognition results are used as a demonstration of this new speaker normalization procedure. Results obtained from using speaker‐adapted templates (the new method), speaker‐independent clustered templates, and speaker‐trained templates are compared.