類神經網路訓練結合環境群集及專家混合系統於強健性語音辨識

近年來，類神經網路（Neural Network）在語音辨識上的研究有著豐碩的成果，有效地減少環境以及語者變異對語音訊號造成的影響，大幅提升辨識率，但系統的語音辨識能力仍有改善空間。本論文即提出新的自動語音辨識系統架構，結合Environment Clustering（EC）、Mixture of Experts與類神經網路以進一步提升系統效能。我們將辨識系統分為Offline與Online兩階段：Offline階段依據聲學特性將整個訓練資料集分割成多個子訓練資料集，並建立各子訓練資料集的類神經網路（以類神經子網路稱之）。Online階段則使用GMM-gate來控制類神經子網路的輸出。新提出的系統架構保留子訓練資料集的聲學特性，強健語音辨識系統。實驗上，我們使用Aurora 2連續數字語音資料庫，依據字錯誤率（word error rate, WER）比較我們提出的語音辨識系統架構與傳統以類神經網路建立的辨識系統，平均字錯誤率進步5.9%，由5.25%降低至4.94%。

Recently, automatic speech recognition (ASR) using neural network (NN) based acoustic model (AM) has achieved significant improvements. However, the mismatch (including speaker and speaking environment) of training and testing conditions still confines the applicability of ASR. This paper proposes a novel approach that combines the environment clustering (EC) and mixture of experts (MOE) algorithms (thus the proposed approach is termed EC-MOE) to enhance the robustness of ASR against mismatches. In the offline phase, we split the entire training set into several subsets, with each subset characterizing a specific speaker and speaking environment. Then, we use each subset of training data to prepare an NN-based AM. In the online phase, we use a Gaussian mixture model (GMM)-gate to determine the optimal output from the multiple NN-based AMs to render the final recognition results. We evaluated the proposed EC-MOE approach on the Aurora 2 continuous digital speech recognition task. Comparing to the baseline system, where only a single NN-based AM is used for recognition, the proposed approach achieves a clear word error rate (WER) reduction of 5.9 % (5.25% to 4.94%).