Triphone Clustering in the Arm System

The use of triphones to cope with contextual effects in phoneme-HMM based speech recognition results in a huge increase in the number of parameters which must be estimated. One solution to this problem is to apply clustering techniques to the triphone set to produce a smaller set of generalized triphones. An alternative is to use knowledge from phonetics of key factors which lead to context-sensitive HMMs. This paper reports an investigation of these methods in the context of the ARM continuous speech recognition system. Experiments confirm that the size of the triphone set can be substantially reduced by clustering with no degradation in recognition accuracy. These results are compared with the outcome of experiments using two knowledge-drive approaches. It is shown that, in this case, superior performance is obtained using the data-driven methods. rh