A fast, sequential decoding algorithm with application to speaker verification

To implement speaker verification (SV) technology for real-world applications with a large user population, the system cost becomes an important issue. One needs a fast algorithm which can support more users in a central telephone switch given the limited hardware, or can reduce the hardware requirement on a wireless handset. Li (see Proc. Int´l. Conf. on Spoken Language Processing, Sydney, 1998) proposed a fast, sequential decoding algorithm for left-to-right HMM. The algorithm is based on a sequential detection scheme which is asymptotically optimal in the sense of detecting a possible change in distribution as reliably and quickly as possible. In this paper, the algorithm is evaluated in a fixed-phrase SV system on a database with 23,578 utterances recorded from 100 speakers. The experimental results show that the decoding speed of the proposed algorithm is about 7 to 10 times faster than the Viterbi algorithm while the accuracy is in an acceptable level. The results indicate that the proposed algorithm can also be applied to speaker identification, utterance verification, audio segmentation, voice/silence detection and many other applications