In this thesis work, we develop a real time continuous digit recognition system using a programmable digital signal processor (DSP). The system uses word-level hidden Markov models (HMMs) for digit vocabularies and the level building algorithm for continuous digit recognition.
First, we extract a 16th order mel-cepstrum feature vector from continuous speech signal through telephone line for every 10 msec frame. The extracted feature vector is then vector quantized and the quantized symbols are used as inputs to HMM. Continuous digit tokens are manually segmented into individual digit sounds and the HMM of each digit unit is constructed using those sounds. For continuous speech recognition, the best path in HMM is found by the Viterbi algorithm within each word, and then the level building algorithm is used to find the best global path for continuous digits. An application system for voice dialing is implemented using a DSP board with a TMS320C31. As a result, we obtain he recognition rate of 88.8% for isolated digits and 66.5% for continuous digits.
