中文摘要：

首先针对现有丢失数据语音识别技术中的边缘化（marginalisation）技术在特征运用上的局限，提出了一种倒谱特征分量的可靠性估计方法，将边缘化技术推广到常用的倒谱语音识别系统中；然后利用基于全带和子带倒谱特征的边缘化识别器在不同噪声中的互补性能，提出了一种噪声自适应的多数据流复合子带语音识别方法。实验结果表明，所提识别方法可以自适应地选出全带和子带数据流中受噪声影响较小者并以之为主要依据进行识别，有效地提高了识别系统在多变噪声环境中的鲁棒性。

英文摘要：

This paper first proposes a new method for evaluating the reliability of cepstral components and extends the marginalisation technique to cepstral recognizers. Then a noise adaptive multi-stream hybrid sub-band approach is proposed for robust speech recognition by making use of the complemental performances between full-band and sub-band cepstral marginalisation recognizers in different noises. Experimental results show that the proposed approach can turn to the less distorted data stream automatically and improve the robustness of the speech recognizer in various noisy environments effectively.