Humaneness-Voice-Large-Alpha-Continuation-Listening

# Humaneness Voice Large: additional 20/40/40 RL listening outputs 70 exact matched scripted takes from the separately identified continuation experiment. The released 1.4 Alpha adapter was continued for 2,000 additional successful updates, retaining AdamW moments with a new warm-up/cosine schedule: 20% Content Enjoyment, 40% matching two fixed supported source-audio emotion heads and 40% genuineness. Sixteen candidates per group, two reference/no-reference cohorts, 32 GPUs. This is not an exact optimizer resume or a replacement of the 1.4 release. The frozen base is laion/Humaneness-Voice-Large-1.3-Alpha. The evaluation retains the exact previous 70 scripts, English directing prompts, seeds, reference codec arrays and frame limits. One take per case; no best-of-N, quality selection, time-stretching or trimming. MP3s are individually streamable, 128 kbps mono. predictions.tar packs MP3s, all twelve MOSS-v2 indices per frame and provenance. metrics.json retains full-clip EARS Medium 192-head/window scores and Parakeet ASR. Fixed prompt-derived evaluation emotion targets are approximate scene-level intentions, not source-audio ground truth and NOT the RL training emotion targets. All outputs, including poor outcomes, remain included. Scores are learned proxies, not listener ratings or a certified held-out benchmark. No speaker-similarity claim. The real/synthetic training mix and reference-source attribution are documented in the linked model and previous public listening datasets. References are not republished here. New generated audio and metadata: CC BY 4.0, LAION, 2026. Listening comparison: https://laion-humaneness-voice-large-alpha-comparisons.static.hf.space/