Development of high-accuracy system for recognizing spontaneous speech

在我们的研究中，我们的目标是提高系统的性能，识别自发语音，这被认为是比识别阅读语音更困难。我们集中讨论了三个技术问题：（1）声学和语言模型，（2）系统组合技术，（3）说话人索引。为了提高声学模型的性能，我们研究了一种基于区分训练的离散混合隐马尔可夫模型，说话人类模型，五音子模型和混响类模型。研究了连续模型与离散模型相结合、多种五音子相结合、混响级模型相结合等系统组合技术。针对语言模型的问题，提出了交叉自适应和交叉验证自适应技术。此外，我们还改进了基于说话人自适应过程中所需的说话人向量的说话人索引技术的性能。

In our research, we aimed to improve the system performance for recognizing spontaneousspeech, which was considered to be more difficult than recognizing read speech. We focused on three technical issues: (1) acoustic and language models, (2) system combinationtechniques, and (3) speaker indexing. For improving the performance of acoustic models,we investigated a discrete-mixture hidden Markov model based on discriminative training, speaker-class model, quinphone, and a reverberation-class model. Some systemco(a) mbinationtechniquesw(a) ere investigated, such as the combination of continuous anddiscrete models, the combination of various quinphones, and the combination of reverberation-class models. For the issues of language models, we proposed the cross adaptation and cross-validation adaptation techniques. In addition, we improved theperformance of speaker indexing techniques based on speaker vectors required during theexecution of speaker adaptation.