Humaneness-Voice-Large-1.2-Alpha-Demo-Audio

# Humaneness Voice Large 1.2 Alpha: listening audio Seven newly written fictional acting scenes in five languages, with two conditions per case: prompt only, or prompt plus a distinct speaker reference. There are 70 selected predictions and five input reference clips. Players: https://laion-humaneness-voice-large-1-2-alpha-listening.static.hf.space/ Model and full training/inference documentation: https://huggingface.co/laion/Humaneness-Voice-Large-1.2-Alpha The target is 34.8 seconds per scene. Actual durations, seeds, full prompts and selection diagnostics are in scenes.json/predictions.json. If the first take was outside 30–40 seconds, up to two additional deterministic seeds were tried; the first duration-valid take is shown, otherwise the closest-duration take. This is duration-only curation, not MOS/WER/reward selection. All attempts, including unselected takes, are retained in all_attempts.tar. No clip is trimmed or time-stretched to disguise a duration miss. audio/ contains individual 128 kbps mono MP3s for browser streaming. selected_audio.tar contains the same selected audio plus metadata and 12-codebook MOSS-v2 arrays. all_attempts.tar includes every candidate with its codec array and JSON. Twelve indices per 80 ms frame = 150 indices/s. The five references are codec-reconstructed crops from the public synthetic Gemini corpus laion/german-english-voice-acting-challenges, CC BY 4.0. references.json retains source UID, provider speaker, duration and attribution. References are inputs, not model predictions or original lossless recordings. The other scripts and predictions are newly authored release artifacts. CC BY 4.0: attribute LAION contributors, Humaneness Voice Large 1.2 Alpha, 2026, and the linked source corpus for references. Respect upstream terms and personality rights. These examples are an experimental listening demo, not a statistically representative quality benchmark or an impersonation tool.