feat[cv2-5]: SPADE distillation training + evaluation on CosyVoice2
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 0
- Forks
- 0
- PR merge metrics
- No merged PRs in 30d
Description
Goal
Iteration 5 of SPADE-on-CosyVoice2: fine-tune the pruned 12-layer LLM with the SPADE composite loss (CE + Skew-KL logits + latent/attention/embedding MSE, dynamic layer matching) using the frozen 24-layer teacher on a real-speech subset, then evaluate teacher/pruned/distilled WER, RTF, and parameters.
Acceptance
-
spade_cosyvoice2/distill.pytrains the pruned student against the frozen teacher with the SPADE loss and saves a CosyVoice-loadable checkpoint -
spade_cosyvoice2/evaluate.pyreports WER (Whisper), RTF, params, depth - Distilled model WER recovers substantially vs the pruned model (verified on real data)
Notes
Iteration 5 for: SPADE applied to CosyVoice 2 (arXiv:2509.20802)
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reading spade_cosyvoice2/distill.py and spade_cosyvoice2/evaluate.py, then trace how the pruned 12-layer student and frozen 24-layer teacher are loaded. Implement the specified SPADE loss and evaluation outputs, and verify completion with a CosyVoice-loadable checkpoint plus real-data WER, RTF, parameter, and depth comparisons.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- audio-video-rtc, machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100