Veda TTS — LJSpeech (CGN v2, 206M)
Fine-tuned from LibriTTS base (ckpt-15000) on LJSpeech.
Model
|
|
| Architecture |
CGN v2 (autoregressive) |
| Parameters |
206M (1024d / 16L) |
| Audio codec |
SNAC @ 24kHz (3-level, 4096 codebook) |
| Text frontend |
Flite G2P (ARPAbet) |
Training
- Dataset: LJSpeech — 12,445 train / 655 val
- Precision: bf16 | LR: 5e-5 (cosine) | Batch: 32
- Best step: 2500 | Early stopped: step 5000
Eval (8 synthesis sentences)
| Metric |
Value |
| eval_loss |
2.727 |
| WER |
5.2% (Whisper base.en) |
| DNSMOS |
3.24 |
Demo
🎙️ Try it
