Humaneness-Voice-Large-1.3-Alpha-RL-Listening

# Humaneness Voice Large1.3 Alpha: matched final-RL listening comparison 70new outputs from the completed rank128 RL adapter on the published final Large1.3 SFT checkpoint (step9,799). These are NOT smoke or intermediate outputs. 35existing scene/language cases, each with and without the same separate input speaker reference. The prior1.2/1.3SFT outputs are unchanged in their original repositories. All words, directing prompts, seeds, requested frames and reference codec arrays match the existing public root-page comparison exactly. One fixed take per condition. No best-of-N, retries, audio editing or time-stretch. Public inference keeps the baseline deployment sampler: BF16 base, unmerged FP32 adapter residuals, temperature1, top-p0.95, top-k50. RL training used FP32 sampling and replay at temperature0.7 with full categorical support. The difference is intentional and documented, not hidden behind a claim of identical policies. The reward is33% Humaneness EARS Medium Content Enjoyment/10,33% matching TWO fixed source-audio emotion intensities (1−MSE after scaling by4), and34% genuineness/6. VESPO sequence weights and separate within-group reward-component normalization are used, with KL/SFT anchoring. G8, separate reference/no-reference groups for each UID; equal real/Gemini and balanced five-language source strata. These are learned proxies, not human MOS or a guarantee of improved output. audio/ contains individually streamable128kbps mono MP3s. predictions.tar packs those MP3s, all12MOSS-v2 codec indices per80ms frame and full JSON provenance. TASKS.json and predictions.json retain exact inputs, original seeds, checkpoint identity, adapter SHA256, requested/actual timing and truncation flags. metrics.json retains full192-head EARS outputs and multilingual Parakeet ASR for both the old 1.3SFT and newRL MP3s. The unfiltered summary includes every take. Clips longer than30s use balanced non-overlapping judge windows, covering the full clip, with duration-weighted raw-head averages and all per-window scores retained. No tail is discarded; this annotation aggregation does not edit the MP3. Windowed proxy scores need not equal a hypothetical whole-clip rating.1.2is not rescored here. This is a scripted listening/development comparison, not a certified held-out benchmark. No ground-truth emotion-intensity match or certified speaker identity claim is made on these scenes. Speaker references are the same separate public synthetic Gemini utterances already attributed in the prior listening dataset, ultimately from laion/german-english-voice-acting-challenges. References are not republished here. Newly generated audio and metadata: CC BY4.0, LAION,2026. The adapter/checkpoint remains a separately identified experimental RL result; this publication does not replace the SFT base model or relabel old SFT audio. Base model: https://huggingface.co/laion/Humaneness-Voice-Large-1.3-Alpha Listening Space: https://laion-humaneness-voice-large-1-2-alpha-listening.static.hf.space/ Completed attempted groups: 1500. Successful optimizer updates: 1452.