Automatic Speech Corpus Construction from Broadcasting Speech Databases

摘要

The speech corpus often needs to be constructed frequently for the diversified speech synthesis. This paper discusses our efforts on construction of speech corpus automatically from broadcasting speech databases for trainable Text-To-Speech (TTS) system. We present a new framework of automatic speech corpus construction from broadcasting speech databases. We select the clean speech audios from the broadcasting audios with a music detector which is based on speech/music discrimination. An automatic speech sentence segmentation system is used to generate the sentence database from the clean speech audios. At last, a text corpus construction method selects appropriate sentences speech which is maximizing the coverage of the sentence databaseȁ9;s diphones. Experiments show that our method can generate a good speech corpus rapidly with minimum manual intervention.