Miso TTS 8B Documentation

If you prefer using pip, create a Python 3.10 virtual environment, activate it, and install the package in editable mode before running the example script.

python3.10 -m venv .venv
source .venv/bin/activate
pip install -e .
python run_misotts.py

To set up the environment using uv, install uv first, clone the repository, and sync the environment with Python 3.10. You can then run the provided example conversation script which downloads the model from Hugging Face automatically.

# Install uv if not present
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone and setup
git clone https://github.com/MisoLabsAI/MisoTTS.git
cd MisoTTS
uv sync --python 3.10
source .venv/bin/activate

# Run the example
uv run python run_misotts.py

Miso TTS can condition generation on prior audio (voice cloning) by providing a list of Segment objects to the context parameter in the .generate() method. Each Segment requires a speaker ID, the text transcript of the audio, and the audio tensor itself.

import torchaudio

from generator import Segment, load_miso_8b

generator = load_miso_8b(device="cuda")

prompt_audio, sample_rate = torchaudio.load("prompt.wav")
prompt_audio = torchaudio.functional.resample(
    prompt_audio.squeeze(0),
    orig_freq=sample_rate,
    new_freq=generator.sample_rate,
)

context = [
    Segment(
        speaker=0,
        text="This is the transcript for the prompt audio.",
        audio=prompt_audio,
    )
]

audio = generator.generate(
    text="This is the next sentence to synthesize.",
    speaker=0,
    context=context,
    max_audio_length_ms=10_000,
)

Use load_miso_8b to initialize the generator and .generate() to synthesize audio from text. The model currently supports English only. The output is a tensor that can be saved using torchaudio.

import torch
import torchaudio

from generator import load_miso_8b

device = "cuda" if torch.cuda.is_available() else "cpu"

generator = load_miso_8b(
    device=device,
    model_path_or_repo_id="MisoLabs/MisoTTS",
)

audio = generator.generate(
    text="Hello from Miso.",
    speaker=0,
    context=[],
    max_audio_length_ms=10_000,
)

torchaudio.save("miso.wav", audio.unsqueeze(0).cpu(), generator.sample_rate)

Miso TTS 8B is a large model (~8.2B parameters). For interactive use, a high-VRAM GPU is recommended. GPU inference defaults to torch.bfloat16.

VRAM Requirements (Approximate):

• bfloat16/fp16: ~16 GB weights, 24 GB VRAM recommended (e.g., RTX 3090/4090, A5000, L4).
• float32: ~33 GB weights, 40 GB+ VRAM recommended (e.g., A100 40GB, A6000, H100).

Other Requirements:

• CPU: Slow inference. Requires ~20 GB RAM for bfloat16 and ~40 GB for float32.
• Disk: ~30–40 GB free space for model weights, Mimi codec, SilentCipher watermarker, and Llama 3.2 tokenizer.