基於隱藏式條件隨機域聲學模型之強健式訓練演算法

Robust Training Algorithm for Noisy Speech Recognition with Acoustic Modeling of Hidden Conditional Random Field

李秋芬(Chiou-Fen Li)

指導教授 ： 洪維廷

元智大學/電機通訊學院/通訊工程學系/碩士(2009年)

摘要

本論文提出以強健式演算法(Robust Environment-effects Suppression Training，簡稱REST)訓練隱藏式條件隨機域(Hidden Conditional Random Fields，簡稱HCRF)聲學模型，提高HCRF模型對雜訊環境辨識效能，再以鑑別式法則訓練HCRF模型，提高HCRF模型對音節的鑑別力。以混合雜訊的訓練語料訓練出HCRF與HMM聲學模型，在人聲雜訊干擾下，HCRF模型與HMM模型辨識能力比較，HCRF模型錯誤改善率可達18.1%；在有通道效應的測試語料MOSTK辨識結果，HCRF模型比HMM模型有21.2% 錯誤改善率。本論文之主要貢獻為下列三項：
1. 提出基於HCRF聲學模型雜訊補償機制。
2.? 運用REST演算法來提高HCRF模型對雜訊環境的辨識效能。
3.? 經D-REST演算法訓練出的HCRF模型，同時兼具抗背景雜訊與增加音節鑑別力。

關鍵字

抗雜訊 ； 隱藏式條件隨機域聲學模型 ； 語音辨識 ； 強健式訓練

並列摘要

In coordination with the robust training and discriminative training technique, a novel algorithm is proposed in this thesis for generating a set of compact hidden conditional random fields (HCRF)-based acoustic models. Among the related issues and techniques we explore are:

1. Derive the compensation operations with HCRF-based models for noise and channel bias distorted conditions.

2. Apply the robust training algorithm for noisy speech recognition with HCRF-based models.

3. Apply the discriminative training technique with HCRF-based models under multi-conditions training database for adverse speech recognition.

並列關鍵字

HMM ； HCRF ； noisy recognition ； Robust Training Algorithm

參考文獻

[1] L. R. Rabiner, “A tutorial on hidden Markov models and selected applications in speech recognition,” Proceedings of the IEEE on speech recognition, vol. 77, no. 2, pp. 257-286, 1989.

[2] A. P. Varga, R. K. Moore, “Hidden Markov model decomposition of speech and noise,” in Proc. IEEE Int. Conf. Acoustics, Speech, Signal Processing, pp. 845-848, 1990.

[3] Saeed V. Vaseghi, Ben P. Milner, “Noise Compensation Methods for Hidden Markov Model Speech Recognition in Adverse Environments,” IEEE Trans. Speech and Audio Processing, vol. 5, no. 1, 1997.

[5] Wei-Tyng Hong, Sin-Horng Chen, “A robust training algorithm for adverse speech recognition,” Speech Communication, vol. 30, no. 4, pp. 273-293, 2000.

[6] Masaki Ida, Satoshi Nakamura, “Rapid environment adaptation method based on HMM composition with prior noise GMM and multi-SNR models for noisy speech recognition,” IEICE Transactions on Information and Systems, Pt.2 (Japanese Edition), vol. J86, no. 2, pp. 195-203, 2003.