A Study on Improvements of Noise Rebustness for HMM-based Speech Recognition

为了实现鲁棒的连续密度隐马尔可夫模型（Continuous Density Hidden Markov Models，HMM）用于带噪语音识别，本文基于人耳听觉特性，提出了一种对高信噪比频率区域共振峰敏感的频率加权HMM。在该HMM中，高斯概率密度函数的协方差矩阵被固定为频率加权矩阵的逆，以利用群延迟谱的鲁棒性，并将其在频域中的相对感知重要性纳入HMM中。几个频率加权函数和频率加权矩阵的缩放方法使用NOISEX-92的国际数据库进行了检查。单词识别测试的结果总结如下。(1)从每个均值向量得到的平滑功率谱给出了最鲁棒的HMM。(2)将加权矩阵转换为协方差矩阵的最佳缩放使得加权系数的和等于1，或者转换后的协方差的行列式比初始Hynth的行列式大50至150倍。(3)在频率加权HMM中，需要更多的状态来获得鲁棒性。(4)自适应预加重提高了对高频能量较小的噪声的鲁棒性。(5)对于白色、粉红色和汽车噪声，频率加权HMM比标准对角HMM获得6到12 dB的SNR增益。(6)即使用标准的谱减法降噪方法对含噪语音进行预处理，在极低信噪比条件下，频率加权HMM的识别分数也比传统HMM高出约10%。

In order to realize robust continuous density Hidden Markov Models (HMM) for noisy speech recognition, this study develops a frequency-weighted HMM based on the human auditory characteristics which is seseitive to formant peaks in high SNR frequency region. In this HMM,the covariance matrices of Gaussian probability density functions are fixed to the inverse of frequency weighting matrices in order to utilize the robustness of group delay spectra and also to incorporate their relative perceptual importance in frequency domain into HMM.Several frequency weighting functions and the scaling methods of frequency weighting matrices are examined using the international data base of NOISEX-92. The results of word recognition tests are summarized as follows.(1) The smoothed power spectrum derived from each mean vector gives the most robust HMM.(2) The optimum scaling to convert the weighting matrices to the covariance matrices is such that the sum of weighting coefficients is equal to one or the determinants of the converted covariances are 50 to 150 times larger than those of initial HMMs.(3) A larger number of states is required to attain the robustness in the frequency-weighted HMM.(4) Adaptive preemphasis improves the robustness to noises which have less energy in the high frequency region.(5) The frequency-weighted HMM attains SNR gains of 6 to 12 dB over a standard diagonal HMM for white, pink, and car noises.(6) Even when preprocessing the noisy speech by the standard noise reduction method of spectral subtraction, the frequency weighted HMM attains about 10% higher recognition scores in very low SNR condition than the conventional HMM.