セグメンタルNMFによる音声分析合成における基底行列圧縮の検討

摘要

本稿では，セグメント単位の非負値行列因子分解（セグメンタルNMF）を用いた音声分析合成符号化システムを目的とし，セグメンタルNMF によって生成される基底行列をより簡潔に表現する圧縮方法を提案する．非負値制約を用いて行列を分解し，特徴抽出を行う非負値行列因子分解手法を各音素の特徴に基づいてクラスタ化したSTRAIGHT のスペクトル包絡系列に適用し，基底行列を求める．符号化を考えた場合に，従来のSTRAIGHTのスペクトル包絡系列における各基底の次元数が513 次でありメモリ量の大きさが問題となる．STRAIGHT スペクトル包絡系列の基底行列は，非常にスパースであるので，少数のガウス混合分布により近似表現について検討した．%We have proposed an analysis-synthesis method by using non-negative matrix factorization in phoneme-wise speech segments (segmental NMF). In this paper, a concise representation of a matrix of basis functions of STRAIGHT spectral envelopes is proposed to reduce the number of dimensions of the NMF basis functions. Non-negative matrix factorization technique decomposes a matrix of STRAIGHT spectrum envelope sequences into a basis matrix and a weighting matrix, both of which have non-negative elements. The conventional STRAIGHT power spectrum has 513 dimensions, which requires large amount of memory spaces. According to our previous research, the obtained basis matrix for each segmental STRAIGHT spectrum has very sparse components, and it is expected that such a matrix can be represented by simple combinations of some Gaussian mixture distributions. Some preliminary results of the approximation of the basis functions are discussed.