Work: A new approach to the continuous speech recognition. Moscow Institute of Physics and Technology.

Main results: The main idea is to avoid in the recognition procedure the hidden Markov models (HMMs) using and to exploit various decoding techniques. The main problems solved with the help of HMMs are selection of parameters, training the recognition system, and recognition. The problem of selection of parameters we can be formulated as a problem of finding minimal set of phonemes, which are necessary to transcribe the words of given dictionary, under the condition of acceptable discrimination of different words. The discrimination depends on preferable error-correcting capability of the speech code. The error-correcting capability will be evaluated by spectrum of code distances between utterances of the words. The training of the recognition system will be solved by traditional statistical methods. In the problem of recognition, in opposite to the concept based on HMMs, the model of the utterance of a word will be presented in explicit way. This enables us to use sequential decoding and error-correcting methods. As concerns the recognition of continuous speech, optimal methods of construction of the best phrase from available hypotheses of words will be studied. The experimental testing of the approach was carried with the material of isolated, separated and continuous speech with the vocabulary for voice dialing system. The material given consists of the speech of 50 speakers who pronounced about 50000 words.