Automatic Recognition of Synthetic Speech Using an Electronic Model of the Middle and Inner Ear.

摘要

A phoneme-based automatic speech recognition system was developed and tested using synthetic speech. The acoustic signal is divided into short segments for analysis; segments are either a single pitch period of voiced speech or a 10 ms sample of voiceless speech. These segments are independently analyzed and given a phonemic name by three different measures. The sub-phonemic segments are grouped using measures which reflect dynamic changes in the speech signal. Each group of segments represents a phoneme and is identified by simple algorithms operating on the string of phonemically-named segments that form the group. The phoneme-based recognition system was tested using isolated synthesized words which permitted evaluation with connected strings of phonemes but stopped short of requiring development of word boundary rules. The tests consisted of 100 phonemically balanced words containing 281 phonemes. Of these, 245 were correctly identified, 23 were mis-identified, 13 were missed entirely, and 11 were added. However, many of these errors were predictable or understandable and may be overcome at a higher (word or phrase) level. (Author)