15th Annual Conference of the International Speech Communication Association

September 14-18, 2014

On the Role of Missing Data Imputation and NMF Feature Enhancement in Building Synthetic Voices Using Reverberant Speech

Dhananjaya Gowda (1), Heikki Kallasjoki (1), Reima Karhila (1), Cristian Contan (2), Kalle Palomäki (1), Mircea Giurgiu (2), Mikko Kurimo (1)

(1) Aalto University, Finland
(2) Universitatea Tehnică din Cluj-Napoca, Romania

In this paper, we study the role of a recently proposed feature enhancement technique in building HMM-based synthetic voices using reverberant speech data. The feature enhancement technique studied combines the advantages of missing data imputation and non-negative matrix factorization (NMF) based methods in cleaning up the reverberant features. Speaker adaptation of a clean average voice using noisy data is generally better than building a speaker dependent voice using the noisy data. In this paper, we show that the proposed feature enhancement technique can further improve the spectral match between the enhanced feature adapted voice and a clean speaker dependent voice.

Full Paper

Bibliographic reference.  Gowda, Dhananjaya / Kallasjoki, Heikki / Karhila, Reima / Contan, Cristian / Palomäki, Kalle / Giurgiu, Mircea / Kurimo, Mikko (2014): "On the role of missing data imputation and NMF feature enhancement in building synthetic voices using reverberant speech", In INTERSPEECH-2014, 2947-2951.