老人中文語音辨識之初步研究

Abstract

在本論文中，從收集的老人語料建立起一個老人中文語音辨識系統，而這個老人中文語音辨識系統的syllable辨識率達44.72%。然後使用TCC-300聲學模型來進行老人語料的調適，選用的調適方法為最大可能性線性迴歸；並且在特徵參數抽取時，使用聲道長度正規化來改善老人聲音低沉的特性，當老人語料的聲音頻率被彎曲至較相似年輕人時，再作最大可能性線性迴歸的調適。而且重複VTLN加上MLLR的調適方法來改善辨識率。最後也分析老人語音腔調差異對辨識與調適的影響，並發現腔調差異的影響可由調適過程來改善；而經由VTLN加上MLLR的調適過程，可以得到最終的音節辨識率達51.47%。

In this thesis, to build up an elder Mandarin speech recognizer used the collected elder speech corpus, then that syllable recognition to reach 44.72%; moreover, using Maximal Likelihood Linear Regression to adapt the elder corpus by TCC-300 acoustic model. When extracting speech feature, utilizing Vocal Tract Length Normalization to modify the property of the elder voice is to low. When the speech frequency of the elder corpus is warping to be close to the youth speech frequency, we implement the MLLR adaptation; moreover, to use iteration VTLN+MLLR to improve the recognition. Final, to analyze different elder accent to cause distinct result on adaptation and recognition, then we find the MLLR adaptation can decrease the effect by different accent. The VTLN+MLLR adaptation can improve the syllable recognition to reach 51.47%