A Study on Constructing Various Acoustic Models using Distributed Speech Corpora

为了收集各种环境条件下的语音，对公共交通引导、车载信息检索和公共空间引导进行了语音对话系统的现场测试。基于这三个语料库，开发了声学模型训练数据共享基础架构的原型。在这个系统中，人们可以通过查询说话者的年龄、话语的信噪比和音素频率的分布来搜索特定的语音子集。该系统通过共享HMM声学模型各状态下的高斯混合模型的访问次数、分支次数、和平方和等有效统计量来训练一组HMM模型。此外，为了对话语进行特征化，即不需要显式语音活动检测（VAD），在较宽的信噪比范围内开发了信噪比方法。在训练策略方面，除了对话语集进行最大似然训练外，还研究了一种仅使用统计量的模型自适应方法。通过识别实验验证了基于预存储统计量的自适应方法的有效性，自适应训练的模型准确率几乎等同于混合EM算法。

In order to collect speech utterances made under various environmental conditions, field tests of spoken dialogue systems have been conducted for the public transportation guidance, the in-car information retrieval and the guidance for a public space. Based on the three corpora, a prototype of the data sharing infrastructure for acoustic model training has been developed. In the system, one can search for the particular speech subsets by invoking queries on the age of the speakers, SNR of the utterance and distribution of the phoneme frequency. The system can train a set of HMM's by sharing the efficient statistics, i.e., the visiting count, the branching count, the sum and the square sum, for the Gaussian Mixture pdf's for each state of HMM acoustic models. In addition, in order to characterize the utterance, a blind, i.e., does not require the explicit voice activity detection (VAD), method for SNR is developed for wide range of the SNR.As for the training strategy, not only the maximum likelihood (ML) training over the set of utterances, but also a model adaptation method using only statistics has been also studied. The effectiveness of the adaptation approach using pre-stored statistics for each utterance was confirmed through the recognition experiments where the accuracy of the model trained by the adaptation is almost equivalent to the pooled EM algorithm.