利用语音学知识和神经网络识别非特定人的复合元音

Speaker-Independent compound vowels recognition using Neutral Network and phonetic knowledge

本文主要研究：以音素为识别基元，运用语音学知识，对非特定人的普通话复合元音进行识别，其特点是音素识别由神经网络（NN）完成．为了便于利用语音学知识，NN输入层的刺激采用语音的功率谱．用单元音训练的NN识别非特定人的普通话复合元音，识别率是54％，而运用语音学知识后，其识别率提高到90％。

A strategy that utilizes a neutral netwrk and is based on phonemes for the recognition of speaker-independent Chinese compond vowels is discussed. The stimuli to NN input layer are the speech power spectra, and the outputs of NN is are presentation of phonemes. A NN trained by monophthongs is used to recognize the compound vowels.Since the phonetic knowledge is used to analyze the results of initial recognition of the compound vowels, the accuracy of recognbo the compound vowels is improved from 54 % to 90%.