Introduction

LAION just published Humaneness Voice Large 1.7 on Hugging Face (created 2026-11-04T12:17:52Z): open CC BY 4.0 expressive text-to-speech for English, German, French, Spanish, and Dutch. You write the words and the performance — emotion, delivery, pauses, non-speech reactions — optionally with a clean reference WAV for speaker identity.

This is a new supervised branch from Large 1.3, not Large 1.6 with another adapter. The card ships complete merged weights: Acting High-8654 plus its selected rank-32 acting adapter at update 44,250 / scale 1.0. Ordinary inference needs no adapter attach. It is also a different product shape from the NotebookLM Edition podcast drop already on artrealmai.com.

What Shipped

  • Vendor / product: Humaneness Voice Large 1.7 — LAION
  • Modality: generative expressive TTS (voice-acting directions + optional identity reference)
  • Open vs closed: fine-tuned weights CC BY 4.0; MOSS Audio Tokenizer v2 and other deps keep upstream terms; custom code needs trust_remote_code
  • Runnable today: hf download + root infer.py on a CUDA GPU (tested PyTorch / Transformers stack on the card)
  • Concrete fact: ~4.13B FP32 parameters across three safetensors shards (~16.5 GB on disk for the merged root); merged acting adapter at update44,250 @ scale1.0; languages en / de / fr / es / nl
  • Listen surfaces: matched checkpoint comparison, five-language acting showcase, eval set laion/humaneness-voice-gemini-r32-cfg-eval

Run It

hf download laion/Humaneness-Voice-Large-1.7 --local-dir humaneness-large17
cd humaneness-large17
pip install -r requirements.txt
python infer.py --language en --seconds 9.2 --seed 1701 \
  --text 'I thought you had forgotten. You remembered every little detail. I am so glad you came back.' \
  --prompt-file examples/prompt-en.txt --output example.wav

Add --reference-wav clean-speaker.wav for identity conditioning. Supply at least 5.04 s of clean single-speaker speech with different words from the target line; the helper crops to 14.96 s before encoding (the old 2.96 s cap is gone). The loader keeps BF16 compute while the root checkpoint is stored in FP32. The codec is pulled separately from OpenMOSS-Team/MOSS-Audio-Tokenizer-v2.

Prompting lives in PROMPTING.md. Older 1.5/1.6 manuals are useful directing ideas only — their specialist adapter + CFG tips were measured on different weights; do not stack those adapters onto 1.7 just because ranks match.

Limits Worth Knowing

  • CUDA path is the reviewed surface; CPU-only is not validated on the card.
  • Showcase progress is incomplete while ~2,200 conditions finish — a pending case is not a successful take.
  • Reference audio improves identity cues; it is not a guarantee of exact voice lock.
  • Use only voices you are authorized to use. LAION opposes deceptive impersonation; that is guidance, not an extra CC BY restriction.
  • Distinct from site coverage of Humaneness Small, Voice Converter, and NotebookLM Edition — this is the Large acting / expressive merged 1.7 branch.

Original Source

Conclusion

If you want open Large-class expressive TTS with acting directions already merged — five European languages, one download, no adapter ritual for day-one inference — Humaneness Voice Large 1.7 is the November 4 card to pull.

—Anabel ♡