| Presentation # | 6 |
| Session: | Deep Learning for Speech Synthesis |
| Location: | Kallirhoe Hall |
| Session Time: | Tuesday, December 18, 14:00 - 17:00 |
| Presentation Time: | Tuesday, December 18, 14:00 - 17:00 |
| Presentation: |
Invited talk, Discussion, Oral presentation, Poster session
|
| Topic: |
Speech recognition and synthesis: |
| Paper Title: |
SYNTHETIC-TO-NATURAL SPEECH WAVEFORM CONVERSION USING CYCLE-CONSISTENT ADVERSARIAL NETWORKS |
| Authors: |
Kou Tanaka, Takuhiro Kaneko, Nobukatsu Hojo, Hirokazu Kameoka, NTT corporation, Japan |
| Abstract: |
We propose a learning-based filter that allows us to directly modify a waveform of synthetic speech to that of natural speech. Speech-processing systems using a vocoder framework such as statistical parametric speech synthesis and voice conversion are convenient especially for a limited data because it is possible to represent and process interpretable acoustic features over a compact space, such as the fundamental frequency and mel-cepstrum. However, the well-known problem leading the quality degradation of generated speech is an over-smoothing effect lacking some detailed structure of generated/converted acoustic features. To address this issue, we propose a synthetic-to-natural speech waveform conversion using cycle-consistent adversarial networks which doesn't require any explicit assumption about speech waveform in adversarial learning. In contrast to current techniques, since our modification is performed at the waveform level, we expect that the proposed method also enables to generate ``vocoder less'' sounding speech even if the input speech is synthesized by using the vocoder framework. The experimental results demonstrate that our proposed method achieves to 1) alleviate the over-smoothing effect of the acoustic features nevertheless the direct modification method for the waveform and 2) dramatically improve the naturalness of the generated speech sounds. |