laion/Humaneness-Voice-1.1-Alpha-Demo-Audio

Humaneness Voice 1.1 Alpha: matched listening audio

Separate 128-kbps mono MP3 predictions/reference excerpts for the Small/Large Alpha comparison. manifest.json pins literal words, exact prompts, seeds, reference UIDs/crops and speaker conditions. results.json holds take-level provenance, ASR hypotheses and scores; summary.json contains all completed aggregates. This audio-only demo dataset does not publish the local generated codec arrays. Locally the pipeline keeps those arrays, MP3s and scores in a handful of transactional SQLite containers, not loose files.

Parakeet v3 WER, Orange 128-D timbre cosine and frozen Humaneness Ears Medium's 192 predictions are retained once scoring completes. These are model-estimated proxies, not human ratings or verified identities. The main scripts were written for evaluation. Some references come from training material. New Gaston/ElevenLabs complete transcripts were checked against training, but speaker-level independence is not claimed. Exact older Real+Gemini/Ref15 predictions are reused only where weights/input match; no output-dependent take selection or best-of-N is used.

Model weights/newly authored generated examples and declared synthetic references: CC BY 4.0. Upstream/model dependencies and personality rights remain applicable. Obtain consent and do not impersonate people.

Public listening Space Small model Large model

Reinforcement-learning comparison added 14 October 2026

The same 207 cases now also include final rank-128 RL adapters for Small and Large, plus two unchanged-base controls. All four new columns use FP32 arithmetic on the BF16-rounded public Alpha base, TF32 disabled, codec sampling temperature 0.7, top-p 0.95 and top-k 50. Archived Alpha predictions remain untouched (BF16, temperature 1.0). Compare RL with its matched control to isolate the adapter effect. There is no best-of-N selection or evaluation-take filtering.

All twelve columns additionally have Orange WavLM-ID 250-D speaker identity cosine and VoiceCLAP performance-text/reference-audio similarities. Identity is distinct from the retained 128-D timbre cosine. The style text contains GENERAL plus round SCRIPT directions, never literal spoken text. Identity is absent without a target voice; name-only scores compare with an audition not supplied to the model. Ten direction-free assistant prompts have no VoiceCLAP text target. Means, standard deviations, valid counts and paired RL-minus-control differences are in the listening page and summary.json. All measurements use the listening MP3s.

Training used G8 candidate groups, 50% reference-conditioned and 50% no-reference, separately normalized reward components and a custom VESPO objective. With reference: CE/VoiceCLAP/identity 40/30/30; without reference: CE/VoiceCLAP 60/40. The base and judges stayed frozen. RL adapters are listening comparisons here; their weight files are not stored in this audio dataset.