论文部分内容阅读
为解决语音识别系统实用中的说话人口音快速自适应问题,提出了一种动态说话人选择性训练方法。基于说话人选择性训练方法,采用基于Gauss混合模型似然分数计算的置信测度选择训练用说话人,改变训练用说话人的绝对数目选取方式,提高了选取的效能并拓展了选取标准的推广性。根据各个训练用说话人同被适应说话人的不同似然程度,加权地合成动态说话人选择性训练的语音模型,提高了自适应训练的效果。实验表明:该方法使识别率从80.16%提高到84.12%,相对误识率降低了19.96%,在实用中提高了基线系统的识别性能。
In order to solve the problem of rapid adaptation of speaker accent in speech recognition system, a dynamic speaker selective training method is proposed. Based on the speaker selective training method, the training speaker is selected based on the confidence measure of the likelihood score of Gauss mixture model to change the absolute number of training speaker selection, which improves the efficiency of selection and expands the generalization of the selection criteria . According to the different likelihoods of each speaker in training and the adapted speaker, the speech model of dynamic speaker selective training is weighted to improve the effect of adaptive training. Experiments show that this method improves the recognition rate from 80.16% to 84.12% and the relative misrecognition rate decreases by 19.96%, which improves the recognition performance of the baseline system in practice.