音響的特徴を用いた話し言葉の断片発話単位への分割

摘要

It is common for speech synthesis technology to process each sentence as one single and independent unit. However, in human speech production, it is perhaps unusual to process a long utterance as a single discrete unit, and typically a series of short utterance fragments is produced in such cases. Such a fragmentary short utterance is assumed to be a minimal discourse unit, and it is proposed here that similar chunks should be used as the basic units for speech synthesis in order to speed-up the calculation processing. In this paper, the acoustic features of such utterance fragments is modeled by HMM, and the paper reports on the result of an experimental the segmentation of a natural speech corpus into optimal units for processing as utterance fragments according to the the model. The ESP-C casual conversation speech corpus was used as material for the experiment.%現在までの音声合成技術は文を一単位として処理することが多かった．しかし，実対話において，人間は長い発話文を一度に処理することは稀であり，多くの場合，短い断片的な発話を用いる．このような短い断片的な発話を断片発話とし，音声合成の計算処理単位として用いる事を提案する．本稿では，HMMにより断片発話の音響的特徴をモデル化し，そのモデルを用いた断片発話単位へのセグメンテーション実験を行なった結果を報告する．実験には，トピックフリーの雑談対話音声を収録したESP-Cコーパスを用いた．