VoxMel TTS Technical Report
VoxMel generates speech without a learned tokenizer, using an exactly invertible transform to split mel-spectrograms into coarse structure and residual detail.
The model generates Coarse-Mel autoregressively, then predicts temporal and frequency residuals without feeding either back into its generation history. Trained on 100,000 hours of Emilia data, the 0.6-billion-parameter model reports 1.51% English word error rate and 1.29% Chinese character error rate. GRPO training lowers those figures to 1.15% and 1.17%, while improving speaker similarity on both test sets. A distilled four-step version nearly matches the supervised model’s error rates and retains competitive speaker similarity.
HF Daily Papers' note
The model generates Coarse-Mel autoregressively, then predicts temporal and frequency residuals without feeding either back into its generation history. Trained on 100,000 hours of Emilia data, the 0.6-billion-parameter model reports 1.51% English word error rate and 1.29% Chinese character error rate. GRPO training lowers those figures to 1.15% and 1.17%, while improving speaker similarity on both test sets. A distilled four-step version nearly matches the supervised model’s error rates and retains competitive speaker similarity.
HF Daily Papers' note
score 5