摘要
The contributions of the static and dynamic articulatory information to speech recognition were evaluated, and the recognition approaches by combining the articulatory information with acoustic features were discussed. Articulatory movements were observed by the Electromagnetic Articulographic System for reading speech, and the speech signals were recorded simultaneously. First, we conducted several speech recognition experiments by using articulatory features alone, consisting of a number of specific articulatory channels, to evaluate the contribution of each observation point on articulators. Then, the displacement information of articulatory data were combined with acoustic features directly and adopted in speech recognition. The results show that articulatory information provides with additional information for speech recognition which is not encoded in acoustic features. Furthermore, the contribution of the dynamic information of the articulatory data was evaluated by combining them in speech recognition. It is found that the second derivative of articulatory information provided quite larger contribution to speech recognition comparing with the second derivative of acoustical information. At last, the combination methods of articulatory features and acoustic ones were investigated for speech recognition. The basic approach is that the Bayesian Network (BN) is added to each state of HMM, where the articulatory information is represented by the BN as a factor of observed signals during training the model and is marginalized as a hidden variable in recognition stage. Results based on this HMM/BN framework show a better performance than the traditional method.
The contributions of the static and dynamic articulatory information to speech recognition were evaluated, and the recognition approaches by combining the articulatory information with acoustic features were discussed. Articulatory movements were observed by the Electromagnetic Articulographic System for reading speech, and the speech signals were recorded simultaneously. First, we conducted several speech recognition experiments by using articulatory features alone, consisting of a number of specific articulatory channels, to evaluate the contribution of each observation point on articulators. Then, the displacement information of articulatory data were combined with acoustic features directly and adopted in speech recognition. The results show that articulatory information provides with additional information for speech recognition which is not encoded in acoustic features. Furthermore, the contribution of the dynamic information of the articulatory data was evaluated by combining them in speech recognition. It is found that the second derivative of articulatory information provided quite larger contribution to speech recognition comparing with the second derivative of acoustical information. At last, the combination methods of articulatory features and acoustic ones were investigated for speech recognition. The basic approach is that the Bayesian Network (BN) is added to each state of HMM, where the articulatory information is represented by the BN as a factor of observed signals during training the model and is marginalized as a hidden variable in recognition stage. Results based on this HMM/BN framework show a better performance than the traditional method.
