Humaneness Voice Large: complete context manuals

Choose the model you will actually load. These English manuals are designed to fit in an agent context while preserving exact tested directing language, quality gates and evidence boundaries. The interactive Space keeps the larger audio and statistics archive.

1.5 Alpha Original

17,424 Qwen tokens · 10,531 words. Other tokenizers differ. 40 emotions, 15 talking styles and 14 edge cases; original model-specific recipes.

Read formatted manual · Plain Markdown for agent context · JSON lookup registry · Model, README and inference code

1.6 Alpha — merged RealSpeech rank 4 and specialist recommendations

23,703 Qwen tokens · 14,664 words. Other tokenizers differ. 40 emotions, 15 talking styles and 14 edge cases; 72 additional reference-conditioned decisions.

Read formatted manual · Plain Markdown for agent context · JSON lookup registry · Model, README and inference code

What 1.6 means

1.6 = 1.5 plus final RealSpeech rank-4 adapter at scale 1, merged once. There is no new training and no active specialist in the default weights. Do not attach the rank-128 RL or rank-4 ancestor again. Nine specialist routes qualify, 45 use manual-only settings, and 18 retain standard fallbacks. Specialist access remains gated; these are native model-specific adapters, not PEFT.

Earlier rank-4 prompting audio used a separately loaded adapter on a BF16 base; its recipes are inherited guidance. The later matched specialist comparison used the exact FP32-merged parent released as 1.6. Historical audio has not been relabeled as newly generated output. Specialist gains without reference audio are unvalidated.

Agent workflow

Read the matching Markdown once, choose the exact recipe and recorded CFG, and use the supplied recipe compiler. Keep spoken text separate from English GENERAL/SCRIPT directions. Use clean references of 5.04–14.96 seconds, and record model revision, reference crop, seed, frame budget, precision and any new specialist scale. Failed and diagnostic routes are labeled; they are not concealed recommendations.

Full synthesis and listening evidence · Matched specialist comparisons

New contributions CC BY 4.0. Preserve upstream attribution and existing gated specialist terms. No additional rights over reference audio are granted.