音声は話者の声道形状の特性や音響機器の特性などの非言語的特徴によって変形するが，この非言語性の変形に対して凡そ不変な音声の構造的表象が提案されている．これは音声の物理的実体を捨象し，その音響空間内での相対的な動きのみを捉えた物理表象である．また，この構造的表象に基づく音声合成の枠組みが提案されている．この枠組みでは音声を発話内容（語形）と発話者の身体性とに分離して捉え，語形に対して発話者の身体性を付与する（戻す）形で音声を生成しており，幼児の音声模倣をモデル化したものといえる．本研究では，この音声合成系の対象を複数の言語に拡張し，任意の話者性と任意の言語性を独立に処理できる音声合成系を検討する．即ち，語学教師によって発声された日英ニケ国語の音声を網羅する構造的表象に対して，日本語母語話者の身体性を日本語を通して付与することで，未修得言語の語形を音声として実体化することを検討する．この合成系による合成音声の評価を行うとともに，異なる条件下で合成した音声同士の比較を行い，合成手法の改善についても考察する．%Speech acoustics easily vary due to non-linguistic factors such as speaker differences and microphone differences. The authors already proposed a structural representation of speech, where these variations can be effectively removed. This representation discards absolute and static properties of speech events and captures only their relative and dynamic features. Recently, a new framework of speech synthesis based on this structural representation has been proposed. Here, an utterance is characterized by two separate attributes, speaker-independent speech form and speaker-dependent embodiment features found in that utterance. On this framework, a new utterance is generated by realizing a given speech form acoustically and this realization is enabled by providing speaker-dependent embodiment features. This generation process can be viewed as implementation of infants' vocal imitation on a machine. In this report, based on this structural representation, cross-speaker and cross-language voice conversion is implemented. A speaker's two utterances of Japanese and English are modeled as a single speech form. By providing another speaker's Japanese utterance of the same content, the English utterance of that speaker will be generated using the speech form. In this report, the performance of the proposed method is evaluated and its problems arernalso made clear.
展开▼
