This paper describes a speaker independent (Hebrew) phoneme recognition system from continuous speech. The system is based on a heuristic algorithm which performs sequential recognition of phonemes in several paths. Each path involves sequential identification of a string of phonemes by adding one phoneme at each step, to the accumulated phoneme string. The added phoneme is selected to maximize the probability of the new string, taking into account phoneme duration and neighborhood probabilities. The proposed algorithm has several advantages. It incorporates a priori knowledge on phoneme duration and neighborhood, it provides a set of the N best strings (with their probabilities) rather than the best string only and it may easily be implemented by a parallel processor. In an experimental evaluation of the proposed algorithm the following recognition results were achieved : 67.14% correct, 30.00% substitutions, 2.86% deletions and 33.25% insertions. The Viterbi algorithm [1] under the same conditions lead to the following results : 53.64% correct, 34.94% substitution, 11.43% deletions and 22.21% insertions. These results were estimated from the identified strings of phonemes by the well known weighted Levenshtein distance.

Keywords: Continuous speech, Phoneme recognition, Heuristic algorithm.