An Evaluation of Postfiltering for Deep Learning Based Speech Synthesis with Limited Data
Beata Loorincz, Maria Nuțu, Adriana Stan, Mircea Giurgiu · 2020
Recently, deep neural network (DNN) based speech synthesis achieved close to human speech quality and became the state-of-the art in the field of text-to-speech (TTS) synthesis systems. However, a major part of its efficiency comes from the use of large quantity of high-quality speech recordings. When this data is not available, other approaches are still preferred.This paper evaluates the DNN-based postfiltering of the synthesised speech as a means to increase the quality of DNN-based TTS systems trained on very limited speech resources. 20 different systems are compared objectively using the Mel Cepstral Distortion measure. The systems differ in terms of: training data, network architecture, and training method. Out of the 20 initial systems, 7 are evaluated subjectively in listening tests performed for two different speakers. Results show that even when starting from as little as 5 minutes of speech recordings, the postfiltering process improves the quality of the synthetic speech output. So it can, therefore, be used as a training strategy for TTS systems where sufficient high-quality data is not available.