并行子带HMM最大后验概率自适应非线性类估计算法

目前,自动语音识别(ASR)系统在实验室环境下获得了较高的识别率,但是在实际环境中,由于受到背景噪声和传输信道的影响,系统的识别性能急剧恶化.本文以听觉试验为基础,提出一种新的独立子带并行最大后验概率的非线性类估计算法,用以提高识别系统的鲁棒性.本算法利用多种噪声和识别内容功率谱差异,以及噪声在不同频带上对HMM影响的不同,采用多层感知机(MLP)对噪声环境下最大后验概率进行非线性映射,以减少识别系统由于环境不匹配而导致的识别性能下降.实验表明:该算法性能明显优于最大后验线性回归算法和Sangita提出的子带语音识别算法.

Adaptive nonlinear class estimation algorithm using parallel subband HMM maximum a posteriori probability

Although auto speech recognition (ASR) systems have achieved high performance under well-defined conditions, the performance of speech recognition system is deteriorated dramatically with the influence of the background noise and the transmission channel in practical environments. Based on the auditory tests, this paper proposes a new nonlinear class estimation algorithm based on the maximum a posteriori probability of the independent subband to improve the robust of recognition systems. The algorithm utilizes the spectrum difference of noise and recognition objects, and noise with different effects to HMMs in different frequency bands, to reduce the deterioration of recognition performance for the environment mismatch through nonlinear mapping maximum a posteriori probability in noise environments with MLP. The test shows that the proposed algorithm is efficient and outperforms other algorithms, such as maximum a posteriori linear regression algorithm and the subband recognition algorithm of Sangita.