In real-life applications, the speech recognition system errors are mainly due to inadequate detection of speech segments, unreliable rejection of out-of-vocabulary (OOV) words, and noise and transmission channel effects. In this paper, we present the results of several experiments carried out on field vs. laboratory databases and on databases collected over PSN and GSM networks. The main sources of errors are analyzed. Preprocessing techniques as well as HMM adaptation techniques are used to increase the robustness to mismatches between training and testing conditions. We show that a blind equalization scheme improves significantly the recognition accuracy on both field and GSM data. Bayesian adaptation of hidden Markov models (HMM) parameters produces robust models to field conditions. The obtained results prove that HMM adaptation and preprocessing techniques can be advantageously combined, in order to improve ASR robustness.
展开▼
