SPEAKER ADAPTATION USING CONSTRAINED ESTIMATION OF GAUSSIAN MIXTURES

A recent trend in automatic speech recognition systems is the use of c ontinuous mixture-density hidden Markov models (HMM's), Despite the go od recognition performance that these systems achieve on average in la rge vocabulary applications, there is a large variability in performan ce across speakers, Performance degrades dramatically when the user is radically different from the training population, A popular technique that can improve the performance and robustness of a speech recogniti on system is adapting speech models to the speaker, and more generally to the channel and the task, In continuous mixture-density HMM's the number of component densities is typically very large, and it may not be feasible to acquire a sufficient amount of adaptation data for robu st maximum-likelihood estimates, To solve this problem, we propose a c onstrained estimation technique for Gaussian mixture densities, The al gorithm is evaluated on the large-vocabulary Wall Street Journal corpu s for both native and nonnative speakers of American English, For nonn ative speakers, the recognition error rate is approximately halved wit h only a small amount of adaptation data, and it approaches the speake r-independent accuracy achieved for native speakers, For native speake rs, the recognition performance after adaptation improves to the accur acy of speaker-dependent systems that use six times as much training d ata.