A FREQUENCY WARPING APPROACH TO SPEAKER NORMALIZATION

In an effort to reduce the degradation in speech recognition performan ce caused by variation in vocal tract shape among speakers, a frequenc y warping approach to speaker normalization is investigated, A set of low complexity, maximum likelihood based frequency warping procedures have been applied to speaker normalization for a telephone based conne cted digit recognition task. This paper presents an efficient means fo r estimating a linear frequency warping factor and a simple mechanism for implementing frequency warping by modifying the filterbank in mel- frequency cepstrum feature analysis, An experimental study comparing t hese techniques to other well-known techniques for reducing variabilit y is described, The results have shown that frequency warping is consi stently able to reduce word error rate by 20% even for very short utte rances.