Computer Recognition of Connected Speech

A system for obtaining a phonemic transcription from a connected speech sample entered into the computer by a microphone and an analog-to-digital converter is described. A feature-extraction program divides the speech utterance into segments approximately corresponding to phonemes, determines pitch periods of those segments where pitch analysis is appropriate, and computes a list of parameters for each segment. A classification program assigns a phoneme-group label (vowellike segment, fricativelike segment, etc.) to each segment, determines whether a segment should be classified as a phoneme or whether it represents a phoneme boundary between two phonemes, and then assigns phoneme label to each segment that is not rejected as being a phoneme boundary. About 30 utterances of 1–2 sec duration were analyzed using the above programs on an interconnected IBM 7090-PDP1 system. Correct identification of many vowel and consonantal phonemes was achieved for a single speaker using the same speech material that was used for developing the recognition procedure. The time for analysis of each utterance was about 40 times real time.