摘要： PMC技术在提高语音识别的鲁棒性方面发挥重要作用。但PMC技术仍存在一些难点：如何获得精确的卷积噪声模型；如何在低信噪比情况下提高识别性能。该文提出了PMC技术的改进方法：引入新的残差噪声模型以及伪干净语音模型，结合信号增强技术，能有效地提高系统的鲁棒性。实验是在英国剑桥大学的HTK语音识别工具包的基础上进行，结果表明，新的PMC技术在噪声环境下能显著提高识别性能。

关键词: PMC, 加性噪声, 卷积噪声, 残差噪声模型, 伪干净语音模型

Abstract: The model adaptation scheme PMC plays an important role in robust speech recognition. However, There are still some open problems in the PMC: how to get the model of convolutional noise, how to work efficiently in low SNR, etc. In this paper, the innovations are developed in the PMC to overcome these shortcomings in some extents. The new residual noise model and pseudo clean speech models are introduced. Incorporating with the noise-reducing schemes, the new PMC is efficient in improving the recognition accuracy in adverse environments. The experiments have been done based on the Cambridge's HTK toolkit to implement the continuous mandarin digit recognition in noisy environments.

中图分类号: 

• TN912.3