自發性國語語音辨識

Abstract

自發性語音是最接近人們在自然情況下的語音，較能實際應用於日常生活中，因此漸趨於重要。本論文將先用國語sub-syllable HMM 建立聲學模型，從錯誤分析知道，Uncertain 與Particle 常與411 syllable 互相辨識，將Uncertain 不加入辨識器，可提高辨識率，Particle 在聲學上，十分相似411 syllable，只差異語意，聲學模型是難以區別。再以Syllable HMM 建立聲學模型，將可提高辨識率 3.3%，使用skip state 改善Deletion 錯誤，改善並不如預期，從錯誤分析知道Deletion 錯誤大部分由Syllable Contraction 造成。

The spontaneous speech is most close to the speech in natural cases of people and can relatively apply to daily life actually, so become more and more important gradually. The first sets up acoustics model with mandarin sub-syllable HMM. From the error analysis, the recognizing device doesn’t often distinguish between Uncertain and 411 syllable, and between Particle and 411 syllable. Do not put Uncertain model into the recognizing device, can improve the recognizing rate. Particle is in acoustics, very similar 411 syllable, and the language purpose of Particle is only different from the language purpose of 411 syllable and this is difficult to distinguish between their acoustics models. The second sets up acoustics model with mandarin sub-syllable HMM, and then can improve 3.3% of recognizing rate, and uses skip state to improve Deletion error, and then that does not improve so well as expectancy. From Deletion error analysis, we know Syllable Contraction causes the most of Deletion error.