说话人自适应训练方法在连续语音识别中的应用

Speaker Adaptive Training in Continuous Speech Recognition

自适应技术在近年来得到越来越多的重视 ,其中应用广泛的包括MAP、MLLR ,该技术利用少量特定人数据就可以调整码本 ,快速地提升识别性能 ,它要求原始的码本有很好的说话人无关性。本文介绍了结合MLLR自适应的说话人自适应训练 (SpeakerAdaptiveTraining ,以下简称SAT)算法 ,这种方法将每个说话人码本视为说话人无关码本经过线性变换的结果 ,在此基础上训练的说话人无关码本更有效剔除了说话人相关信息 ,因此在说话人自适应中时能根据特定数据调整更好地逼近说话人特性 ,从而有更好的性能表现。

More and more attentions have been paid on speaker adaptation in recent speech recognition research, especially on widely used MAP and MLLR. These techniques apply to fast codebook adjustment when only limited amount of training data is available, and they demand original model to be speaker independent. This article introduces MLLR integrated Speaker Adaptive Training (SAT) method, which regards every individual's codebook as the result of linear transformation of speaker independent codebook and trains speaker independent codebook based on such concept. Since speaker-related information is extracted by this means, the trained codebook is more 'speaker independent', so it would perform better in speaker adaptation.