Voice Conversion Challenge 2016
Compare different voice conversion systems and approaches using the same voice data!
Task of the 1st Challenge
The task was speaker conversion, which was a well-known basic task in voice conversion.
- We provided voices of 5 source and 5 target speakers (consisting of both female and male speakers) from fixed corpora as training data. Each speaker uttered the same sentence set consisting of around 150 sentences.
- Using these parallel data sets, voice conversion systems for all speaker-pair combinations (25 speaker-pairs in total) were developed by each participant.
- Other voices of the same source speakers were provided later as test data that consists of around 50 sentences for each speaker. Each participant generated converted voices from them using the developed 25 conversion systems.
- The resulting 25 converted voice sets were evaluated in terms of perceived naturalness and similarity with listening tests.
We focused on 16 kHz speech and signal-to-signal conversion strategies. No transcription was provided and the use of phonetic information given manually was NOT allowed. Participants were free of using additional data (for training purposes) different from the database from which the speakers corpora were provided.
- November 25th: release of training data
- January 8th: release of evaluation data
- January 15th: deadline to submit the converted audio.
- February 29th: notification of results
Participants needed to strictly follow the Challenge rules:
Summary of the Challenge is shown in the following page:
This work was supported in part by
- Tomoki Toda (Nagoya University)
- Linghui Chen (University of Science and Technology of China)
- Daisuke Saito (Tokyo University)
- Fernando Villavicencio (National Institute of Informatics)
- Mirjam Wester (University of Edinburgh)
- Zhizheng Wu (University of Edinburgh)
- Junichi Yamagishi (National Institute of Informatics)
Contact information: vcc2016__at__vc-challenge.org