Experiments in Automatic Speech Recognition

Abstract

Use of the phoneme as a basic unit for automatic speech recognition is attractive from the standpoint of storage requirements and correspondence to conventional phonemic analysis. However, the segmentation of speech into phonemes and the specification of the identifying characteristics of a phoneme have presented difficulties. It has become clear that the physical speech signal does not consist of clearly defined units corresponding to phonemes. Likewise, perception of phonemelike units is influenced by adjacent units. Nevertheless, speech can be segmented with some degree of consistency at points which approximate phoneme boundaries, and these segments can, in many cases, be identified with moderate accuracy This paper describes the results of an automatic-speech-recognition research program oriented toward phoneme recognition. Input speech consisting of CVC monosyllables spoken in isolation was first analyzed with an 18-channel filter bank whose rectified outputs were quantized to 8 levels in 3.5 dB steps and recorded in digital form on magnetic tape. These digital recordings then served as input to an RCA 501 computer. A program for automatic segmentation and identification of phonemes was developed. Automatic vowel location and recognition accuracy averaged 45% for 10 vowels spoken in 60 CVC monosyllables by 10 male talkers, with 10 repetitions by each talker; scores for individual talkers as high as 85% were obtained. No adjustment was made for individual talkers. Initial consonant-identification accuracy averaged 60% for 6 consonants (/s, ∫, z, v, f, h/) with average scores for individual consonants ranging from 30% to 80%, and average scores for individual talkers as high as 75%. [This work was sponsored by the U. S. Air Force Systems Command.]