中文語音辨識中聲韻母混淆集合之研究

Abstract

本論文中嘗試去找出中文語音辨識中﹐聲韻母之混淆集合。我們定了五個聲母混淆集合及六個韻母混淆集合。這五個聲母混淆集合分別為(1)不送氣爆破音及空聲母: /ㄅ, ㄉ, ㄍ, 空聲母/, (2)送氣爆破音及喉擦音: /ㄆ, ㄊ, ㄎ, ㄏ/ (3)鼻音及流音: /ㄇ, ㄋ, ㄌ, ㄖ/ (4)上顎音: /ㄐ, ㄑ, ㄒ/ (5) 捲舌及非捲舌音: /ㄓ, ㄔ, ㄕ, ㄗ, ㄘ, ㄙ/。六個韻母混淆集合分別為 (1)/ㄚ/-集合, (2) /ㄛ-ㄨ/-集合, (3) /ㄜ/-集合, (4)/ㄝ/-集合, (5) /ㄧ/-集合, (6) /ㄩ/-集合。 我們討論了二階梅爾倒頻譜特徵參數及能量參數對聲韻母混淆集合的影響﹐並且改用最小分類錯誤訓練法來訓練模型參數並探討這樣的訓練法則對於聲韻母混淆集合的影響。最後﹐我們提出兩個簡單的方法來解決聲韻母的混淆集合。聲母加強方法可以增加不送氣爆破音及空聲母集合以及送氣爆破音及喉擦音集合的辨識率並且提升整體辨識率。對於韻母混淆集合﹐我們提出一個較為強健的辨識方法來解決因為語者本身唸錯所造成的錯誤。實驗證明此方法可以把 [ㄅㄛ, ㄆㄛ, ㄇㄛ, ㄈㄛ] 的錯誤減少 25% 而 [ㄅㄥ, ㄆㄥ, ㄇㄥ, ㄈㄥ] 的錯誤減少42%。

In this thesis, we attempt to discover the confusing sets in Mandarin speech recognition. We define five INITIAL confusing sets and six FINAL confusing sets. The five INITIAL confusing sets are (1)Unaspirated plosive and Null set: /b, d, g, null/, (2) Aspirated plosive-glottal fricative set: /p, t, k h/ (3) Nasal-Lateral set: /m,n,l,r/ (4) Palatal set: /j, q, x/ (5) Retroflex-Nonretroflex set: /zh, ch, sh, z, c, s/. The six FINAL confusing sets are (1) /a/-set, (2) /o-wu/-set, (3)/ e/-set, (4) /eh/-set, (5) /i/-set, (6) /yu/-set. We examine the effect of delta delta mel-frequency cepstrum features and energy features on INITIAL and FINAL confusing sets. We also discuss the influence of MCE training on INITIAL and FINAL confusing sets. Finally, two simple methods are proposed to solve the INITIAL and FINAL confusing sets. The INITIAL weighting method can improve the unaspirated plosive and null set and aspirated plosive and glottal fricative set. For FINAL confusing sets, we provide a more robust recognition method to solve two particular errors caused by speaker customary error. The experiment results demonstrates that the errors of /b, p, m, f/-o are reduced by 25% and the errors of /b, p, m, f/-eng are reduced by 42%.