基于特征分量输出概率加权的多数据流鲁棒语音识别方法

Robust multi-stream speech recognition based on weighting the output probabilities of feature components

针对传统多数据流语音识别方法不考虑数据流内各特征分量受噪声影响差异的缺点,提出了一种基于特征分量输出概率加权的数据流结合新方法,分析了特征分量输出概率加权对识别的影响,并结合丢失数据技术中的边缘化(Marginalisation)模型和软判决(Soft decision)模型给出了两种具体的数据流结合方案。将所提数据流结合方案应用到复合子带语音识别系统中,实验结果表明,所提识别方法可以根据噪声环境的不同自适应地调整数据流对识别影响的大小,其性能显著优于传统的多数据流识别方法。

Traditional multi-stream fusion methods in speech recognition try to control the stream influences on the decision by weighting the stream outputs. This paper proposes a new stream fusion method which weights not only the stream outputs, but also the output probabilities of feature components. The effect of the new fusion method on stream influences on the decision is discussed and two stream fusion schemes based on the mariginalisation and soft decision models in missing data techniques are also proposed. Experimental results on hybrid sub-band speech recognizer show that the proposed approaches can adjust the stream influences adaptively and outperform the traditional multi-stream methods in various noisy environments.