快速口音自适应的动态说话人选择性训练

Dynamic speaker selected training for rapid speaker adaptation

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。

Practical speech recognition systems need rapid speaker adaptation to be effective with a wide variety of speakers. A dynamic speaker selected training method developed for rapid speaker adaptation improves the basic speaker selected training method by replacing the absolute number selection method used in the basic method with a confidence measure calculated from the Gaussian mixture model likelihood. The new method enhances both the training speaker selecting efficiency and the selecting adaptability. The dynamic acoustic model, which uses different weightings for each training speaker so that they resemble the adapted speaker, further increases the recognition accuracy rate. Simulation show that the dynamic method improves the baseline recognition accuracy rate from 80.1% to 84.1%, with a decrease of 19.96% in the relative error rate. Thus, the dynamic method rapidly increases practical speech recognition system performance.