以類神經網路為基礎之中文連續語音辨認系統

Abstract

本論文主要探討以遞迴式類神經網路為架構之中文連續語音辨認系統。並初步嘗試將模組化類神經網路辨認方法推廣至大字彙、不特定語者之辨認，其具體方法，是以一個辨認語者性別的類神經網路模組做性別分群，再進一步也將次模組化類神經網路分成兩組，使訓練運算量減低為原來的一半，在平行處理下能增加訓練以及辨認的速度。而訓練的方法是採用四階段訓練法，分別對次音節、音節、字串階層及性別階層作訓練。實驗結果與同條件下的HMM做比較，連續音節之辨認結果(62.23%)比HMM辨認法(59.55%)要好，顯示我們的嘗試初步成功。

In this thesis we extend the modular recurrent neural network (MRNN) based speech recognition approach to speaker-independent, continuous Mandarin speech recognition. It employs a sophisticated MRNN to attack the complicated task. The MRNN is composed of two gander-dependent sub-MRNNs for the discrimination of 411 base-syllables and a gander classification RNN for combining the outputs of these two sub-MRNNs. Each sub-MRNN can be further divided into three parts: two RNNs for the discriminations of 100 right-final-dependent initials and context-independent 39 finals, two weighting RNNs for the generation of dynamic weighting functions for combining initial and final discriminant functions, and one RNN for the detection of syllable boundaries to provide timing cues for the recognition search. The whole system is trained by a four-level training scheme including sub-syllable-, syllable-, utterance-, and gender-level trainings. Experimental results showed that the proposed method outperformed the conventional HMM method. The base-syllable accuracy rate raised from 59.55% obtained by the HMM method to 63.23% obtained by the proposed MRNN method.