Multi-condition Training and System Combination for Automatic MOS Prediction

The quality of text-to-speech synthesis and voice conversion has been improved. To evaluate their quality, the mean opinion score (MOS) is widely used but its scoring requires human evaluation. Automatic MOS prediction can reduce development costs. The VoiceMOS challenge is a shared task for automatic MOS prediction in which the MOS of synthesized speech taken from the previous Blizzard challenges and voice conversion challenge is predicted. The challenge has main and out-of-domain (ood) tracks; we found that automatic MOS prediction systems show poor performance for the ood track. To improve this, we introduce multi-condition training. Experiments show that multi-condition training significantly improved the MOS prediction performance. Visualizing the embedded vectors of MOS prediction models shows that multi-condition training increased the margin between samples from two tracks.