基于REMOS的远距离语音识别模型补偿方法

封闭环境中远距离语音识别会受到混响效果的影响，从而降低语音识别率。混响建模(reverberation modeling for speech recognition，REMOS)是一种在模型域进行混响补偿的新方法，该方法在已知声源位置的情况下能有效提升远距离语音识别精度。但在实际应用中，往往难以预测声源的位置。利用最大后验概率的原理，基于对房间不同区域进行有区别补偿的思想，在按帧的隐马尔可夫模型 (hidden Markov model，HMM)补偿的基础上，提出一种在封闭环境中新的模型补偿方法。该方法利用K均值聚类K-means算法对房间冲击响应 (room impulse response，RIR)的优化集进行聚类，对所属相同类的混响模型进行合并处理，再把合并后的混响模型载入维特比算法中，对清晰语音的HMM模型进行按帧补偿。最后采用后验概率方法选择最佳补偿，使得模型域的混响补偿能最接近精确补偿。实验证明，该方法能进一步提升远距离语音识别的精度。

REMOS-based method for model domain compensation in remote speech recognition

The distant-talking speech recognition would be affected by reverb in a enclosed environment. As a result, the recognition rate would be greatly reduced. Reverberation modeling for speech recognition(REMOS) is a new method for reverberate compensation in the model domain; it can improve distant-talking speech recognition accuracy effectively if the sound source location is already known. But in a real application, location of sound source can be hardly to predicted. Based on the principle of maximum a posteriori probability and frame-wise hidden Markov model(HMM) model compensation, a new method for model compensation in a enclosed environment is proposed in this paper. In this method, K-means clustering algorithm is used to cluster room impulse response(RIR) optimized sets, and merge the reverberation model which is in a same kind class, then Viterbi decoding algorithm is loaded, and frame-wise compensation is implemented to the clear speech HMM model. At last, the best compensate model is selected through the maximum a posteriori estimation. It makes model domain reverberate compensation to be closest to the accurate compensation. The experimental results prove that the method can enhance distant-talking speech recognition accuracy further.