Bayesian feature enhancement for large vocabulary speech recognition in the presence of noise and reverberation

该项目的目标是开发一个大词汇量的连续语音识别系统，该系统对噪声和混响具有鲁棒性，如果语音被远距离麦克风捕获，这是典型的失真。为了保证开发的解决方案的广泛适用性，假设只有单通道记录的可用性。调查的出发点是一方面的贝叶斯特征增强方法，已开发的一个合作伙伴，并已被证明是非常有效的小识别任务。另一方面是另一个项目合作伙伴的大词汇量连续语音识别（LVCSR）系统，该系统已成功用于许多国际项目和基准测试。贝叶斯特征增强算法将进一步发展，以满足更高的要求，一个大的词汇量的任务。进一步的相互作用的功能增强与Sophistated LVCSR系统必须进行调查的最佳集成，以实现一个强大的大词汇识别系统的遥远的语音。

The goal of this project is the development of a large vocabulary continuous speech recognition system that is robust towards noise and reverberation, which is the typical kind of distortion if the speech is captured by distant microphones. In order to guarantee a wide applicability of the developed solutions, the availability of only single-channel recordings is assumed. The starting point of the investigations is on the one hand a Bayesian feature enhancement method that has been developed by one partner and which has been shown to be very effective on small recognition tasks. On the other hand there is the large vocabulary continuous speech recognition (LVCSR) system of the other project partner, which has been used successfully in many international projects and benchmarks. The Bayesian feature enhancement algorithm will be further developed to meet the higher requirements of a large vocabulary task. Further the interaction of the feature enhancement with the sophistated LVCSR system has to be investigated for an optimal integration, in order to realize a powerful large vocabulary recognition system for distant speech.