Researchers have developed a text-to-speech system for Mizo, a low-resource tonal language spoken primarily in the Indian state of Mizoram, using only 5.18 hours of training data. They compared a baseline Tacotron2 model with a VITS model, both trained on the same data, to see whether an end-to-end non-autoregressive approach could handle tone without explicit tone markings.
In objective and subjective evaluations, VITS outperformed Tacotron2 on Mel-Cepstral Distortion and listener ratings, and it also produced numerically lower tone errors. However, differences in DNSMOS, F0 RMSE, and F0 correlation were not statistically significant, so the advantage was not uniform across all metrics.
The authors conclude that a non-autoregressive, end-to-end framework can achieve acceptable perceptual quality and intelligibility for a low-resource tonal language, even without explicit tone markings. This suggests that such models may be a practical path for expanding TTS coverage to under-resourced languages.