Introduction

LAION just shipped Humaneness Voice Large — NotebookLM Edition on Hugging Face: open CC BY 4.0 weights for a two-person podcast text-to-speech stack that generates the whole exchange as one acoustic sequence. You give it a written dialogue and two clean 5–15 second voice references, and it speaks the conversation in one pass instead of synthesizing each turn and joining files.

The card went up late on November 2, 2026 (UTC) and was last touched early November 3. Primary artifact: laion/Humaneness-Voice-Large-NotebookLM-Edition. There is also a listening Space, humaneness-voice-large-podcast-listening.

This is a different drop from Humaneness Voice Small (Sep 29) and the later Voice Converter. Large is the NotebookLM-style multi-speaker podcast path.

What Shipped

  • Model: Humaneness Voice Large — NotebookLM Edition by LAION.
  • Size: about 4.13 billion BF16 parameters (sharded safetensors; ~8.5 GB on disk with assets).
  • Job: two-speaker podcast TTS from dialogue JSON + two identity reference WAVs.
  • Generation style: one continuous acoustic take for the requested exchange — not per-turn concat.
  • Audio: MOSS Audio Tokenizer v2 decodes to 48 kHz mono; twelve codebooks every 80 ms frame.
  • Architecture: Large MOSS Local Transformer — 36-layer Qwen3-family semantic transformer (2,560-d) plus a one-layer local acoustic transformer at the same width.
  • Lineage: starts from Humaneness Voice Large 1.3 Alpha, gets a three-epoch full fine-tune on filtered two-speaker podcasts, then a rank-64 quality-ladder fine-tune already merged at scale 1. Download root weights and run; do not apply the included adapter a second time.
  • License: fine-tuned weights CC BY 4.0; tokenizer, upstream code, and source recordings keep their own terms. Custom code needs trust_remote_code.

Languages and Quality Notes

LAION is blunt about language strength on new text in their HQ demo study:

  • English mean per-clip ASR WER about 1.3% (n=9)
  • German about 6.3% (n=9)
  • Spanish mixed, about 16% (n=6)
  • French treated as unreliable (~95% WER) — stress condition, not support

Across 30 HQ listening examples, mean WER was about 24%, with short clips cleaner than long ones. Metrics are model/ASR estimates on training-domain identities with new text, not a held-out-speaker contest. Listen before you ship anything.

How to Run It

You need a CUDA GPU with enough memory, Python 3.11+, and a recent PyTorch. Their tested stack used PyTorch 2.9.1 and Transformers 5.14.1. CPU-only and other precision modes are not validated on the card.

hf download laion/Humaneness-Voice-Large-NotebookLM-Edition --local-dir notebooklm-edition
cd notebooklm-edition
pip install -r requirements.txt
python podcast_infer.py --model . --dialogue examples/dialogue.json \
  --reference-s1 examples/en-S1.wav --reference-s2 examples/en-S2.wav \
  --seed 20261103 --output conversation.wav

The first run also pulls the separately licensed MOSS Audio Tokenizer v2. Swap in your own authorized clean single-speaker WAVs for S1 and S2. Each reference must land between about 5.04 and 14.96 seconds (longer files are cropped). Optional earlier conversation context can add up to 60 seconds.

Pro Tip

Start with two short turns from examples/dialogue.json, then stretch. Durations and speaker switches are learned instructions, not hard cuts — sampling uses audio temperature 1, top-p 0.95, top-k 50, CFG 1, so whole-dialogue takes stay stochastic. Read PODCAST_PROMPTING.md before inventing your own ChatML shape.

Where It Fits

If you want NotebookLM-flavored two-host shows with open weights you can audit, this is the clearest same-day path: script in, refs in, one WAV out. It is research-grade — LAION asks you to use only voices you have rights to, skip deceptive impersonation, and test language and speaker stability before production.

For tiny single-language CPU voices instead, see recent Dispatches like TiHon-TTS and PiVo-1.4M. For the earlier small acting ladder, stay with Humaneness Voice Small.

Conclusion

A merged ~4B open podcast TTS that speaks a whole two-person dialogue in one acoustic pass is a real step past turn-by-turn glue. Pull the card, try the English or German examples, and hear whether continuous dialogue generation is the voice workflow you have been waiting for.

—Anabel ♡