Skip to content
Audio

Scylla's Band v4

речь и пение в одной лёгкой модели

2 min readupdated

Scylla's Band v4-20261106 от acidsound — локальная TTS-модель, которая синтезирует речь и пение в 14 голосах на 7 языках, включая корейский. Одна flow-модель рендерит оба режима, а декодер выбирается по типу доставки: Vocos для речи и BigVGAN v2 для пения. Есть runtime на C ABI 3.1.0 для Linux, Windows, macOS, iOS, visionOS и Android.

Ключевые особенности

  • Речь и пение из одного flow: пение по JSON-партитуре с текстом, нотами и долями, стили pop/ballad, 14 голосов.
  • Корейский: нормализация чисел, дат и валют, G2P по стандартному произношению, 87-фонемный vocab без переобучения.
  • Быстрый локальный инференс: C++ runtime на CPU даёт 18.9–42.9× реального времени (RTF 0.02–0.05); 4 Euler steps.
  • Мультиплатформенность: LiteRT, Core ML, Core AI, ONNX Runtime и standalone WebGPU-демо в браузере.
  • Проверка: TLA+ 34 276 состояний без нарушений; CTest 16/16 и Pytest 131/131; паритет C++/Python 0.0.
  • Управление: intensity, pace и pitch от −2 до 2; для фраз менее 4 фонем intensity/pace остаются голосовыми.

Пример

python -m scyllasband download --flavor litert
python -m scyllasband speak --voice scylla --language ko -o hello_ko.wav '안녕하세요, 오늘 날씨가 정말 좋네요.'
python -m scyllasband sing --voice scylla --language ko --score melodies/phrase_ko_twinkle.json -o twinkle_ko.wav
python serve_webgpu_demo.py --host 0.0.0.0 --port 8000
ПодробнееПроизводительность: на MacBook Pro M5 Max LiteRT CPU — ~34×, Core ML GPU — ~170×; iPhone 17 Pro Max Core ML GPU — ~65×.
Качество: CER 0.0% на ключевых корейских фразах по faster-whisper large-v3-turbo.
Ограничения: модель не поёт свободно — нужна JSON-партитура; управление глобальное, без кривых и эмоций; до 512 фонем на граф-предложение.
Железо: локальный CPU/GPU, Apple Core ML/Core AI, LiteRT, ONNX Runtime и WebGPU; отдельных требований к VRAM в карточке нет.
Discuss Telegram 0 82

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

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.

    # acidsound/scyllasband
    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 acidsound/scyllasband --local-dir ~/models/scyllasband
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to scyllasband
  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 scyllasband (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

acidsound/scyllasband
Author
acidsound
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

820

Ask AI about this model

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

Comments