Skip to content
Audio

caramel-1

речь в речь без единого слова текста

1 min readupdated

caramel-1 — экспериментальная speech-to-speech модель от naklecha на 413M параметров, вдохновлённая Dessa. Она обучена только на записях человеческой речи и никогда не видела текст; код, рецепт обучения и веса открыты.

Ключевые особенности

  • Две копии модели общаются друг с другом в реальном времени: каждые 80 мс каждая решает, говорить или молчать.
  • Генерация примерно в 11 раз быстрее реального времени на GPU; на MacBook Pro M3 Pro две копии вместе идут примерно в 4 раза быстрее реального времени.
  • Нет voice activity detector и правил очерёдности: модель предсказывает момент речи так же, как аудио.
  • Обучена на 11 555 часах публичной речи: 5 743 ч диалогов и 5 812 ч монологов; за обучение прошла около 33B токенов.
  • Лицензия Apache 2.0; есть скрипты запуска и обучения.

Пример

./setup.sh ./caramel-1
source ./.venv/bin/activate
python generate.py --model ./caramel-1 --seconds 30 --realtime
python generate.py --model ./caramel-1 --prompt opening.wav --seconds 30 --out talk.wav
ПодробнееUTMOS 2.49 против 2.50 у реальных людей; одновременная речь 8.8% против 9.0%.
Обучение заняло около 2 ч 15 мин (~4 GPU-часа).
Ограничение: многие фразы пока не имеют смысла, а детектор всё ещё легко отличает синтез от людей (3.49 против −0.92).
Лицензия Apache 2.0.
Видео-пример
Discuss Telegram 15 155

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.8 GB · ready files1.3 GB
  • FP8 / INT8recommendedweights 0.4 GB0.9 GB
  • 4-bitweights 0.2 GB0.7 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 2060

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.

    # naklecha/caramel-1
    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 naklecha/caramel-1 --local-dir ~/models/caramel-1
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to caramel-1
  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

New posts about versions of caramel-1 (quants, finetunes, LoRAs) will show up under the bell.

Fine-tunes 3All on Hugging Face

About the model

naklecha/caramel-1
Author
naklecha
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

15515

Ask AI about this model

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

Comments