A speaker-independent, syntax-directed, connected word recognition system based on hidden Markov models and level building

In the last several years, a wide variety of techniques have been developed which make practical the implementation and development of large networks for recognizing connected sequences of words. Included among these techniques are efficient and accurate speech modeling methods (e.g., vector quantization, hidden Markov models) and efficient, optimal network search procedures (i.e., level building). In this paper we show how to integrate these techniques to give a speaker-independent, syntax-directed, connected word recognition system which requires only a modest amount of computation, and whose performance is comparable to that of previous recognizers requiring an order of magnitude more computation. In particular, the recognizer we studied was an airlines information and reservation system using a 129 word vocabulary, and a deterministic syntax (grammar) with 144 states, 450 state transitions, and 21 final states, generating more than 6 × 10⁹sentences. An evaluation of the system, using six talkers each speaking 51 test sentences, yielded a sentence accuracy of about 75 percent resulting from a word accuracy of about 93 percent, for an average speaking rate of about 210 words per minute.