X2-NativeCursor: Native-Token Text Progress Tracking for Incremental-Text Streaming Codec TTS
Paper • 2609.09677 • Published
X2-NativeCursor-CosyVoice2-0.5B
X2-NativeCursor observer head trained for the native-token interface of CosyVoice2-0.5B. This repository contains the observer only, not the CosyVoice2 base model.
Checkpoint
- File:
head.pt - Selected seed: 0, selected by minimum
dev_partialMAE - Parameters: 2,892,653
- Native frame: 40 ms
- SHA-256:
a77dc495ccd450c69a1260cc577538fba264f57cd5e5094ac82335f681ce0d9d
Results
Chinese test subset: 600 utterances, with 2,000-sample utterance bootstrap confidence intervals.
| Seed | Chinese-subset MAE | 95% CI |
|---|---|---|
| 0 (selected) | 0.2843 | 0.2368–0.3426 |
| 1 | 0.2543 | 0.2224–0.2919 |
| 2 | 0.2411 | 0.2104–0.2788 |
Selected seed-0 breakdown:
| Metric | Value |
|---|---|
| Plain MAE | 0.4435 |
| Numeric MAE | 0.2292 |
| Symbolic MAE | 0.1802 |
| All-population MAE | 3.0179 |
| English-containing MAE | 11.2187 |
| P90 nominal lag | 40 ms |
| Onset F1 at 80 ms | 0.7479 |
Download
hf download zehan1/X2-NativeCursor-CosyVoice2-0.5B head.pt
Paper: arXiv:2609.09677
- Downloads last month
- 55