Abstract: The paper introduces a new feature compensation method which will induct the relativity of the prediction of spectrum based probability model in detail. The method evaluates the parameters of the joint distribution using the expectation maximization (EM) algorithm. The minimum mean squared error (MMSE) estimator for the speech feature parameters in spectrum-domain based the prior probability distribution is to enhance the correctness of speech recognition. The algorithm is tested in different noise and signal noise ratio (SNR). Subjective measure testifies that this method can increase the correctness of continuous speech recognition.

Key words: Speech recognition, Denoising, Spectrum difference, Probability model

摘要： 在概率模型中，给出了引入倒谱预测值的动态相关性来进行特征补偿的方法。该方法采用期望最大化(EM)算法来估计联合分布参数，基于语音和噪声的先验概率密度，在倒谱域中对语音特征参数进行最小均方误差预测(MMSE)，以提高语音识别精度。不同噪声环境和不同信噪比下的实验结果表明，该方法能有效地提高噪声环境下的中文连续语音识别的正确率。

关键词: 语音识别, 噪声抑止, 倒谱差分, 概率模型

CLC Number: 

• TP301.6