自然音声の合成への一歩

概要

論文の詳細を見る
This paper describes speech as a three-part process, portraying meaning through the intersection of text and voice. We distinguish the language-specific aspects of a spoken utterance from the speaker-specific characteristics of the voice, and show how these can be integrated, even for multi-lingual utterances, using a basic synthesis engine that remains flexible and free of language or speaker dependencies. We present a method for labelling a corpus of speech so that it can be re-cycled to produce novel utterances reproducing the speaking style and voice characteristics of the recorded speaker. In addition, the paper discusses the use of non-speech sounds that are common in spoken conversation and proposes that these may be an essential part of the perceived naturalness of speech. It suggests that future interfaces using speech synthesis may benefit from the controlled use of such non-speech noises as laughter and audible breaths to ease the interaction with a human user.
一般社団法人情報処理学会の論文
1997-02-07