HuZhouSpeech: A Huzhou Dialect Speech Recognition Corpus

Abstract

In this paper, a new open source dialect speech corpus is proposed. It is by far the only open source Goetian dialect corpus, suitable for conducting speech recognition research and building a speech recognition system for Huzhou dialect. This corpus contains about 184 hours of speech and corresponding subtitle transcriptions. In addition, an automated method for building speech recognition corpus is proposed. Only the film and television materials with subtitles are needed to build the speech recognition corpus automatically. Tested on four classical speech recognition frameworks to demonstrate the effectiveness of the corpus for training different speech recognition systems.