Skip to content
Audio

Humanness Voice Podcast

TTS для многоголосых подкастов

2 min readupdated

Humanness Voice Podcast от LAION — полный fine-tune TTS-модели Humaneness Voice Large 1.7 на шаге 9 874, не LoRA. Это ~4,13 млрд параметров в BF16: архитектура MOSS Local Transformer сочетает Qwen-подобный semantic backbone с локальным acoustic transformer и генерирует 12 acoustic codebook tokens на кадр 80 мс. MOSS-v2 codec отдельно рендерит моно-аудио 48 кГц; модель озвучивает сценарий с краткими указаниями по подаче и фиксированными референсными голосами, а пайплайн собирает из реплик подкаст.

Ключевые особенности

  • Полный fine-tune, не LoRA: BF16-веса на ~4,13 млрд параметров (8,26 ГБ), обучение на шаге 9 874.
  • Reference-conditioned TTS: на каждый голос нужен чистый якорь примерно 5–15 секунд; модель говорит ровно по сценарию.
  • Разрешены короткие inline-указания по подаче — разреженные, спокойные и естественные; стиль одобрен в listening-тесте.
  • Best-of-N на каждый ход: по умолчанию N=4, отбор по genuineness и ECAPA-сходству с референсом; N=1 — один дубль без ранжирования.
  • Пайплайн собирает MP3, субтитры, JSON-таймлайн и listening-страницу; оптимизирован под одну RTX 3090 24 ГБ.

Пример

hf download laion/Humanness-Voice-Podcast --local-dir humanness-podcast
cd humanness-podcast
python3.12 -m venv .venv
source .venv/bin/activate
pip install torch==2.6.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt
python optimized_pipeline/prepare.py
python infer.py --model . --script optimized_pipeline/script-spontaneous-en.json --n 4 --groups 2 --device cuda:0 --output runs/en-best4
ПодробнееНа RTX 3090 24 ГБ: немецкий N=4 — RTF 0.821, WER 1.98%, genuineness 3.943; английский strict N=4 — RTF 1.013, WER 5.50%.
Полные BF16-веса — 8,26 ГБ, но codec, judges, KV-кэши и CUDA graphs требуют дополнительной памяти.
Обучение покрывает английский, немецкий, французский, испанский и нидерландский, но оптимизированный CLI и сравнения сейчас оценивают только английский и немецкий; для обоих рекомендованы одни и те же немецкие референсы.
Качество для остальных языков и голосов не измерено одинаково; strict Best-of-4 может выбрать неполное предложение — в примере это отдельно исправлено.
Discuss Telegram 0 81

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 7.7 GB · ready files9.0 GB
  • FP8 / INT8recommendedweights 3.8 GB4.7 GB
  • 4-bitweights 2.2 GB2.9 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 3080

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # laion/Humanness-Voice-Podcast
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download laion/Humanness-Voice-Podcast --local-dir ~/models/Humanness-Voice-Podcast
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

  1. moss-tts-local-transformer-4.55b-voice-acting-v2family root
  2. moss-tts-local-transformer-4.55b-voice-acting-v2-sft3
  3. Humaneness-Voice-Large-1.2-Alpha
  4. Humaneness-Voice-Large-1.3-Alpha
  5. Humaneness-Voice-Large-1.7
  6. Humanness-Voice-Podcast

New posts about versions of moss-tts-local-transformer-4.55b-voice-acting-v2 (quants, finetunes, LoRAs) will show up under the bell.

Fine-tunes 2All on Hugging Face
LoRAs and adapters 9All on Hugging Face

About the model

laion/Humanness-Voice-Podcast
Author
laion
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

810

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments