用于噪声鲁棒语音识别的声学模型及解码策略

A Novel Acoustic Model for Noise Robust Speech Recognition and its Decoding Tactics

提出了一种平行子状态隐马尔可夫模型用作噪声鲁棒语音识别的声学模型。该模型融合了纯净语音和背景噪声信息,模型的每个状态包含平行关系的子状态。在此基础上,提出了两种用于平行子状态隐马尔可夫模型的识别解码策略——子状态最大似然解码和联合转移子状态最大似然解码。实验结果表明,声学模型及其解码策略在各种噪声下取得了良好鲁棒识别效果。

In this paper, a parallel sub-state hidden Markov model, which integrates the clean speech and noise information, and each state of the model has several parallel sub-states,is presented. Then sub-state maximum likelihood and combining transition sub-state maximum likelihood （CTSSML）for parellel sub-state hidden Markov model are also presented. The experiments show that the parallel model and corresponding decoding tactics achieve excellent robust results under each kind of noise.