Humaneness-Voice-Large-1.3-Alpha

Humaneness Voice Large 1.3 Alpha#

Final checkpoint: step 9,799, completed on 22 October 2026. This repository contains the complete model, not a LoRA or a reinforcement-learning adapter. It generates expressive speech from literal text and English performance directions, optionally with a separate voice reference or a registered fictional voice name. Supported training languages are English, German, French, Dutch and Spanish. This is the Large model, not Humaneness Voice Small.

Listen: matched 1.2 / final 1.3 predictions · Additional five-language emotion-masking study · Prompting guide · Exact training recipe and reproduction · Voice names · Inference code

What changed from 1.2?#

We started from the released Large 1.2 Alpha final checkpoint, step 7,851, and fully fine-tuned both the semantic backbone and the local Talker. All data were mixed throughout one continuous learning-rate schedule. The run combines five-language real speech, the existing Gemini acting data, a new English set with emotion masking and physical-performance situations, French-accent characters, and add-on voices. Eligible new English recordings appear twice; the other eligible targets appear once. There is no separate English-only or Gemini-only final phase, and no LR restart between sources.

All synthetic prompts use the new Detailed Prompts conversion: scene intent is kept in GENERAL, and supported local performance beats are attached to aligned speech chunks in SCRIPT. Literal words, existing duration/pause annotations and vocal-burst instructions are preserved. Uncertain local beat alignment stays global rather than inventing timing. Intended or hidden emotions are not proof that those emotions are audible. Real-speech prompts retain their observed descriptions and timed scripts.

Validation sets differ from 1.2, so the loss values are not a cross-release quality comparison. The listening page provides matched prompts, references and seeds. No new objective CE, WER or speaker-similarity improvement is claimed solely from training loss. Future RL with CE/emotion/genuineness is a separate proposal, not part of these weights.

Quick start: generate a WAV#

Use a CUDA GPU. The tested stack uses PyTorch 2.9+, Transformers 5.14.1, Accelerate 1.14.0 and BF16. Review the Python files before enabling trust_remote_code=True. Download only this complete model and the separate MOSS Audio Tokenizer v2 codec; loading an additional parent model is unnecessary.

python -m pip install 'huggingface_hub>=1.29,<2'
hf download laion/Humaneness-Voice-Large-1.3-Alpha --local-dir ./humaneness-large-1.3
cd humaneness-large-1.3
python -m pip install -r requirements.txt

python infer.py --model . --language en --seconds 6 --seed 20261022 \
  --text "You stayed with me through every doubtful evening. I will never forget that." \
  --prompt 'GENERAL: Warm, grateful and naturally conversational.\nSCRIPT:\n(sincere gratitude) [3.6 seconds duration] You stayed with me through every doubtful evening.\n[0.3 seconds pause]\n(warm, quietly smiling) [2.1 seconds duration] I will never forget that.' \
  --output gratitude.wav

For reference-conditioned speech, add --reference-wav another_utterance.wav. Use clean, single-speaker audio with a different transcript, ideally 8–15 seconds. The helper requires at least 5.04 seconds and caps the encoded crop at 14.96 seconds. It does not silently truncate to the old 2.96 seconds. The reference supplies voice identity; the script supplies the words and requested performance. Identity transfer is approximate, not guaranteed.

For a trained named voice instead, add --voice-name Aurora (or another exact name from the catalog). Do not provide a reference at the same time. Leave both options absent for unanchored prompt-only generation. There are 55 registered fictional aliases; they are not tokens to insert into the spoken text. The helper places them in the exact trained wrapper.

--prompt-file directions.txt accepts a UTF-8 GENERAL/SCRIPT file instead of --prompt. --text always contains only words to speak. Output is 48 kHz mono PCM WAV. --seconds is a conditioning target, not a hard guarantee of exact duration. The helper refuses to overwrite an existing output.

Architecture in plain language#

A Qwen3-family semantic transformer reads the words, directions and any reference audio codes. Its state conditions a small local autoregressive Talker, which predicts the acoustic codebooks for each frame. A separate frozen MOSS Audio Tokenizer v2 decoder converts the codes into sound.

Twelve codebooks do not mean 12 tokens per second. This trained model uses the first 12-codebook representation, not a claim that all 32 codec books were supervised. The supplied helper explicitly encodes/decodes the trained 12-book layout. See config.json and stats/MODEL_PREFLIGHT.json.

Repository contents and verification#

• Root: complete weights, configuration, tokenizer, model/processor code, portable infer.py, voice catalog and dependency pins.
• training/jupiter/: frozen runtime source for the actual 1.3 run, including prompt conversion, data-map preparation, training and Slurm scripts.
• training/legacy/: imported shared training, packing, loss and model helpers.
• stats/: original run/data contracts, loss traces, checkpoint metadata, optimizer-resume smoke proof, BF16 export hashes and public-inference checks.
• evaluation/: exact listening tasks, reference provenance and generation code.

All exported tensors were compared with the final FP32 source after BF16 conversion through the public model loader. Thirteen tied output heads are verified explicitly. The public CLI is additionally tested with reference WAV encoding, a registered alias and prompt-only generation; see stats/PUBLIC_INFERENCE_CANARY.json. The three small canary outputs are packaged together in evaluation/public_cli_canary.tar.

The BF16 export does not include the original AdamW/RNG checkpoint payload. Those complete training states remain in the original run. Starting a new fine-tune from these public weights uses a new optimizer; it is not a bit-for-bit resume. TRAINING.md explains data availability, original paths and the changes required outside JUPITER.

Data, attribution and intended use#

Synthetic acting sources and fictional voice catalog: laion/german-english-voice-acting-challenges. The exact newly annotated five-language real-speech cohort is locally prepared; this model release does not claim that the complete cohort or new prompt sidecars are already mirrored publicly. See the source manifests and training notes for the distinction between raw counts, approved prompts, held-outs and actual training exposure.

Released model and LAION-authored materials: CC BY 4.0. Attribute LAION contributors, Humaneness Voice Large 1.3 Alpha, 2026. The model continues the MOSS Local Transformer / SFT3 lineage. The external codec is OpenMOSS-Team/MOSS-Audio-Tokenizer-v2. Preserve upstream notices and comply with each dependency's and dataset's own terms; this release does not relicense third-party code or recordings.

This is an experimental research release intended for expressive speech research, authorized voice work and dataset augmentation. Test quality, language coverage and identity behavior before deployment. Use only voices and recordings you are authorized to use. We oppose deceptive impersonation, fraud and non-consensual voice cloning. These are use recommendations, not additional restrictions on CC BY 4.0. LAION does not endorse downstream misuse or guarantee fitness, factual correctness or exact identity transfer.