CosyVoice2 vs CosyVoice3 on IMDA NSC FEMALE_01
Download printable cheat-sheet (CC-BY 4.0)07 Feb 2026, 00:00 Z
Experiment Status: Not production-ready (current run)
60-second takeaway
In this benchmark, CosyVoice2 baseline audio sounded acceptable while our CosyVoice3 finetune run did not meet production quality.
This is a run-specific outcome under one configuration, not a broad claim that CosyVoice3 is inherently worse.
We are treating CosyVoice2 as a control baseline and CosyVoice3 as an active rerun candidate.
Where this fits
- For founders: do not deploy this CosyVoice3 run as-is.
- For engineers: use this page as a diagnostic handoff for the next rerun.
Series overview:
Result summary
CosyVoice2 is included as a baseline/control and produced acceptable qualitative output on our selected sample. CosyVoice3 finetuning in this run did not reach production-ready quality, with unstable long-form behavior and weaker linguistic consistency in listening checks.
Audio evidence
CosyVoice2 baseline/control
<audio controls preload="none"> <source src="/assets/audio/tts/cosyvoice2/cosyvoice2_baseline_remotion.wav" type="audio/wav" /> </audio>CosyVoice3 representative sample from this run
<audio controls preload="none"> <source src="/assets/audio/tts/cosyvoice3/cosyvoice3_epoch8_remotion.wav" type="audio/wav" /> </audio>What this does and does not mean
This conclusion is specific to our exact setup: dataset shape, checkpoint path, prompt handling, and inference configuration. It should not be interpreted as a universal model-family ranking.
Likely contributors in this run
- Checkpoint quality drift after early epochs.
- Sensitivity to prompt formatting and long-text decoding behavior.
- Operational fragility from large checkpoint churn and unstable inference zones.
Next rerun plan
- Rebuild a controlled CosyVoice3 run with stricter checkpoint gating and explicit long-text eval checkpoints.
- Keep a fixed prompt formatting harness for all comparisons.
- Re-evaluate using the same rubric and side-by-side audio panel used in this series.
Engineer appendix
Key paths from this run
- CosyVoice LoRA tools added:
/mnt/work/chee-wei-jie/voice-models/CosyVoice/tools/train_cosyvoice3_lora.py,/mnt/work/chee-wei-jie/voice-models/CosyVoice/tools/infer_cosyvoice3_lora.py - CV3 run artifacts referenced in sessions:
