Analysis of acoustic Models Trained On a Large-Scale Japanese Speech Database

摘要

This paper investigates the performance of speaker-independent (SI) acoustic hidden-Markov-models (HMMs) trained with a huge Japanese speech database, and discusses the efficiency and task-independency involved. The database consists of read and spontaneous speech uttered by 3,771 speakers. The speech involves wide distributions with respect to region and age to capture the Japanese speech characteristics as best as possible. Recognition experiemnts using the spontaneous speech show that task-independent acoustic models can be created when training data with a huge number of speakers is available.