Introduction

LAION just shipped a new open podcast TTS product on Hugging Face: laion/Humanness-Voice-Podcast (created 2026-11-07T18:27:45Z). It is a full supervised fine-tune of Humaneness Voice Large 1.7 at update 9,874, trained for scripted multi-speaker episodes — not another packaging of the acting checkpoint, and not the older NotebookLM Edition path.

You write the dialogue with speaker labels and emotion directions, drop in one clean reference recording per voice, and the included pipeline generates candidates, checks the words, trims unused audio, and assembles an episode. It does not write the podcast script for you.

What Shipped

  • Vendor / product: Humanness Voice Podcast — LAION
  • Modality: open-weight multi-speaker podcast text-to-speech (turns or pairs) plus episode assembly
  • Open vs closed: fine-tuned weights CC BY 4.0, ungated full checkpoint (no LoRA merge step); MOSS Audio Tokenizer v2 and upstream deps keep their own terms; custom code needs trust_remote_code
  • Runnable today: hf download laion/Humanness-Voice-Podcast then root infer.py on a CUDA GPU (tested stack on the card: Python 3.13.5, PyTorch 2.9.1, Transformers 5.14.1)
  • Concrete fact: two-epoch SFT on roughly 80% real-podcast / 20% acting replay, 32 GPUs, 9,874 optimizer updates; MOSS Local Transformer with a Qwen-family semantic backbone generating twelve codec channels per 80 ms frame, decoded to 48 kHz mono

Two Generation Paths

  • turns: one speaker turn per call with that person's anchor — stronger identity control, easy to replace a single turn
  • pairs: two consecutive turns together (usually S1 then S2) with both anchors — keeps the local back-and-forth inside one generation

Both paths can batch texts, prompts, speakers and candidate seeds. You can keep one candidate for speed or two (or four / eight) for a small selection pool. Every raw candidate and accept/reject decision lands in a SQLite journal, and a listening HTML page updates as units appear.

Pro Tip

Pin --revision or download a local directory before a long resume. A later documentation commit on the same repo can otherwise resolve to a different snapshot while your run is in flight.

How to Run It

hf download laion/Humanness-Voice-Podcast --local-dir humanness-podcast
cd humanness-podcast
python -m pip install -r requirements.txt

# One referenced speaker per generation; two seeded candidates per turn
python infer.py --model . \
  --script examples/short_en.json --output runs/en-turns \
  --mode turns --candidates 2 --batch-size 2 --compact-pauses

# Generate two turns at once, then assemble the chosen blocks
python infer.py --model . \
  --script examples/episode_de.json --output runs/de-pairs \
  --mode pairs --candidates 2 --batch-size 2 --compact-pauses

Open runs/.../listening.html while it works. The finished show is episode.wav, with episode_timeline.json for the assembly clock. Pass --codec /local/codec if you already keep MOSS Audio Tokenizer v2 on disk. Reference clips should be clean single-speaker audio of about 5–15 seconds (codec window 5.04–14.96 s).

Quality Snapshot and Limits

On LAION's earlier fixed-gap segment/block comparison (ten scripts per condition, automatic metrics from EARS Medium + Parakeet Redux — not a blinded listening study), the Podcast checkpoint at step 9874 improved block intended-identity cosine from 0.429 (parent Large 1.7) to 0.730, with block WER down from 10.2% to 2.2%. Segment WER stayed high; treat those numbers as research diagnostics, not a production SLA.

Languages on the card: English, German, French, Spanish, Dutch. Detailed reference/segment listening so far focuses on English and German. Speaker collapse, late switches, silence/mumbling tails, and approximate durations are still known failure modes. Optional quiet-pause compaction and optional Voice Converter post-processing are marked unvalidated on the card.

Where It Fits

  • Distinct from our NotebookLM Edition coverage (whole-dialogue path from the Large 1.3 line) and from Large 1.7 acting TTS
  • This drop is the 1.7 podcast SFT + turns/pairs pipeline + ASR/identity ranking + episode assembly package makers can pull today
  • Research / dataset-augmentation release: use only voices you have rights to, disclose synthetic speech, and do not use it for deceptive impersonation

Conclusion

If you already liked Humaneness Large 1.7 for acting takes and wanted a same-family open pipeline for two-host shows with candidate QA, Humanness-Voice-Podcast is the zero-day to clone. Script in, two references in, episode WAV out — under CC BY 4.0 weights you can audit.

—Anabel ♡