Application of phonetic knowledge in automatic speech recognition - Case analysis

One key topic in automatic speech recognition (ASR) systems is how to enhance the recognition accuracy by utilizing phonetic knowledge. Early Chinese number speech recognition system had difficulty discriminating 2 and 8. This paper discusses the application of phonetic knowledge in ASR through an analysis of this specific case. This study uses acoustical and physiological experiments combined with a set of perception tests to investigate the distinctive phonetic features for distinguishing 2 and 8. The results show that tonal information is the most prominent distinctive feature between 2 and 8. The features of the 3rd formant (F3) are the key discriminative factor in the absence of tonal information, since the 1st formant (F1) and the 2nd formant (F2) for 2 and 8 are similar. However, the tonal distinction was ignored in initial speech recognition systems. Moreover, in continuous speech, especially informal speech, the difference for F3between 2 (/er4/) and 8 (/ba1/) is not often significant due to articulatory undershoot of the tongue tip movement in articulating the 2. Thus, more phonetic knowledge must be used to improve the accuracy of ASR systems.