关键词: 最大似然线性回归（MLLR）, 最大后验概率（MAP）, 环境自适应, 远场语音识别

Abstract:

Adaptive technology can adjust the acoustic model parameters with less data to achieve better speech recognition effect. Most of them are used to adapt to accented speech. In this paper, Maximum Likelihood Linear Regression（MLLR） and Maximum A Posteriori（MAP） adaptive techniques are used in distant noise reverberation environment to analyze their recognition performance in this background. The experimental results show that under the simulation conditions, the reflection coefficient of the wall is 0.6, and the MAP has the best adaptive performance under various noise environments. When the Signal-to-Noise Ratio（SNR） is 5 dB, 10 dB, 15 dB, respectively, MAP reduces the Word Error Rate（WER） by 1.51%, 12.82% and 2.95% on average. Under real conditions, MAP reduces the WER to a maximum of 37.13%. On this basis, the good graduality of MAP is verified. When the number of adaptive sentences is 1000, the WER of distant noise reverberation continuous speech obtained by MAP acoustic model adaptive method is 12.5% lower than that precious adaptive period.

Key words: Maximum Likelihood Linear Regression（MLLR）, Maximum A Posteriori（MAP）, environmental adaptation, distant speech recognition