Improvement of Very Large Vocabulary Speech Recognition using an encoding based on probabilistic structure of vocabulary

对于语音识别，在大词汇量任务中，如果没有语言模型的贡献，任何音素序列都不会导致统计上显著的不足。另一方面，对于说话人识别/验证来说，除了增加训练数据之外，似乎很难实现。大多数说话人识别应用中，不能期望有足够的训练数据。此外，很难预先假设目标音素序列。因此，需要一种新的方法来进行说话人识别，因为许多以前的语音识别方法不能有效地用于说话人识别。

For speech recognition, in a large vocabulary task, any phone sequence didn't induce statistically significant deficient performance without contribution of language models. For speaker recognition/verification, on the other hand, it seems to be difficult other than increasing of training data. Most applications of speaker recognition, it cannot be expected sufficient training data. Moreover, it is difficult to assume a target phone sequence in advance. Therefore, a new method is required for speaker recognition, because many previous methods for improving speech recognition cannot be efficient for speaker recognition.