In real-life applications, the speech recognition system errorsare mainly due to inadequate detection of speech segments, unreliablerejection of out-of-vocabulary (OOV) words, and noise and transmissionchannel effects. In this paper, we present the results of severalexperiments carried out on field vs. laboratory databases and ondatabases collected over PSN and GSM networks. The main sources oferrors are analyzed. Preprocessing techniques as well as HMM adaptationtechniques are used to increase the robustness to mismatches betweentraining and testing conditions. We show that a blind equalizationscheme improves significantly the recognition accuracy on both field andGSM data. Bayesian adaptation of hidden Markov models (HMM) parametersproduces robust models to field conditions. The obtained results provethat HMM adaptation and preprocessing techniques can be advantageouslycombined, in order to improve ASR robustness
展开▼
