Tree Based Clusteringを利用した音節波形接続型音声合成法に関する検討

摘要

音声合成法の手法の1つとして，音節波形接続型音声合成法が提案されている．この手法の問題点の1つとして，任意の一般名詞を作成する際に大量の録音単語が必要となることが挙げられる．そこで収録されているデータベースに対してTree BaSed Clusteringを行うことで，理論上は全ての音声を作成出来る．しかし，音声品質が非常に悪い音声もある.本論文では，異なる2つのモデルでTree Based Clusteringを行い，音声品質の改善を目指す．音声品質の評価には，オピニオン評価実験および対比較実験を用いる．オピニオン評価実験の結果，標準モデルの音声が1．7，拡張モデルの音声は2．5という値が得られた．また，対比較実験の結果，標準モデルの音声が20％，拡張モデルの音声が80％となった．%Word synthesis by concatenating syllabic components method is proposed as a speech synthesis method. As a problem of this technique, large amount of recording words is needed when we make an arbitrary general noun. Then, all speech ideally can be made by doing tree-based clustering for collected database. However, very low quality speech is generated sometimes. In this paper, our aim is to improvement of the speech quality by doing tree-based clustering with two different models. The mean opinion score (MOS) and the ABX test are used for the evaluation of the speech quality. As a result of the MOS, the standard model was obtained 1.7 and the enhanced model was obtained 2.5. Moreover, as; a result of an ABX test, the standard model was obtained 20% and the enhanced model was obtained 80%.