Sequence-To-One Neural Networks for Japanese Dialect Speech Classification

要旨

Automatic speech recognition (ASR) is usually constructed for recognizing standard language. Thus, when input speech includes dialect which is a variety of a language, performance of ASR is seriously degraded. To relax this problem, an approach is to use dialect-specific ASR recognizers by introducing a dialect speech classification module. In this situation, the performance of dialect-specific ASR depends on that of dialect speech classification. We propose a Japanese dialect speech classification method using sequence-to-one neural networks that are one of the successful methods in speech classification research fields. The experimental results showed that a classification system provided high classification accuracy.

主題