This paper presents a time and memory efficient multistage word candidate hypothesizer suitable for medium-size vocabulary applications on small hardware. It is based on a novel compact speech representation: regions of high phoneme similarity values. The processing stages of the word hypothesizer are applied in sequence to reduce the search space for a more computationally expensive fine match word recognition system. The paper also presents a scoring procedure for combining information from each stage of the hypothesizer with the output of the fine match procedure to produce the final word decision. On a 100 word task, use of the word hypothesizer reduced alignment complexity by 93% (compared to exhaustive search by the fine match alone), with significant error rate reduction for clean and noisy test data due to score combination.
Bibliographic reference. Morin, Philippe / Applebaum, Ted H. (1995): "Word hypothesizer based on reliably detected phoneme similarity regions", In EUROSPEECH-1995, 897-900.