LAION · Chatterbox S3Gen voice conversion
One source. Two voices. Four models.
Open the English-first curated comparison, including untouched vanilla Chatterbox →
The same 160 emotional source clips and the same two target voices as the original S10 listening room. Compare the 300-hour supervised baseline with three 24,000-group RL follow-ups. All four models use the same source/target pair, seed, 99-axis conditioning and ten flow steps per case.
How to read this comparison
The source audio supplies words and dynamic emotion/style. Target A or B supplies speaker identity, its acoustic prefix and static timbre-related scores. Switch the target below; each row then plays the original source and four voice-converted outputs side by side.
The best step for each RL arm was selected from steps 8k, 16k and 24k using 40 source UIDs not used as RL training sources. The comparison below also includes 37 known overlapping source UIDs, clearly marked. Do not treat the in-training reward model as an independent quality judge.
Fixed target voices
Both references are 5–6 seconds long. They come from the linked page; no new target was substituted.
Results at a glance
Values below use only the UID-disjoint cases not used for checkpoint selection. Higher CE/PQ/speaker cosine and lower top-two-emotion MSE are directionally better. CE and emotion are training reward-model proxies, not human ratings.
| Model | Step | CE ↑ | PQ ↑ | Emotion MSE ↓ | Speaker cosine ↑ |
|---|
Methods, caveats and downloads
For each RL arm, an identical 40-case, two-target UID-disjoint selection set chose the checkpoint maximizing 0.4 × CE improvement / baseline SD + 0.4 × emotion-MSE reduction / baseline SD + 0.2 × target-speaker-cosine improvement / baseline SD. Selection is exploratory and can still overfit this small set. The summary table excludes these 40 selection cases and all 37 known RL-training-source overlaps.
Content Enjoyment and emotion scores come from frozen Humaneness Ears Medium; speaker cosine uses CAMPPlus. The 160 source MP3s were re-tokenized from the linked page, so inputs are shared exactly across these four models but may differ slightly from the page's earlier preparation. No ASR/WER or human MOS claim is made here. The original S10 page and this experiment both include training-distribution material; individual listening is essential.
Training weights and instructions: Continue, Re-merged, Speaker-Mix. The 300-hour supervised baseline is in the parent release.