In this thesis work, we develop a keyword spotting system based on discrete hidden Markov model(HMM) and we reduce the performance degradation in telephone environment using cepstral mean normalization(CMN) techniques.
For keyword spotting, the best word sequence is found by frame-synchronous Viterbi search algorithm with word-pair grammar. In order to improve the performance of keyword spotting for telephone environment, we apply the CMN method to cope with the telephone channel distortion and resulting vector quantization error. And we propose the local CMN technique that obtain cepstral mean vectors from adjacent frames in continuous speech to compensate the differences between training phase of isolated words and recognition phase of connected words, and the resulting keyword recognition rates are significantly improved for clean speech and channel distorted speech.
We implement the on-line keyword spotting system on SUN SPARC workstation using audio interface.
