The final Large1.3Alpha full-SFT model (step9,799) remains the frozen base. A new rank128 attention adapter in both the semantic backbone and local Talker was optimized using VESPO sequence importance weights with separately group-normalized reward components, a group-relative policy-gradient method. This is not the flow-matching GROW method used for the separate Voice Converter model.
For each source example, eight candidates were sampled in a reference-conditioned group and eight in a separate no-reference group. Both groups use the same source utterance and are accepted or rejected together. Training uses equal real-speech/Gemini proportions and balanced English, German, French, Dutch and Spanish source examples. Candidate transcripts are checked with multilingual Parakeet ASR; WER above15%, invalid audio or truncation disqualifies a candidate.
| Component | Weight | Meaning |
|---|---|---|
| Content Enjoyment | 33% | Humaneness EARS Medium audiobox:CE, scaled by10. |
| Two-emotion matching | 33% | 1 minus mean squared error to two FIXED strongest supported emotions in the source target audio, after scaling by4. Heads are not reselected from each candidate or from the speaker reference. |
| Genuineness | 34% | Humaneness EARS Medium genuineness_0_6, scaled by6. |
The reporting reward is0.33×CE/10 +0.33×(1−MSE_top2)+0.34×genuineness/6. Policy advantages normalize each component within its group; sampled KL and a gold SFT anchor (each0.02) regularize the update. Speaker similarity and VoiceCLAP are NOT reward terms in this run. Learned scores are useful proxies, not human listener ratings or proof of a quality improvement.
Final recorded progress: 1500 attempted groups; 1452 successful optimizer updates. Rejected groups remain logged. Checkpoint files retain adapter weights, AdamW, all rank RNG states, schedule and sampler position.
Checkpoint adapter SHA256: c26031b9e1932ae0e76b29ec894b49bb6f719a9ff997cf733eae8bbf443f9fb2. The selected checkpoint is the completed final run, not a reward-selected “best” or a smoke checkpoint.