Miso TTS 8B

State-of-the-Art Text-to-Speech Model

Quickstart | Model Introduction | Model Summary | Usage | Safety

---

Quickstart

To quickly try the model, you can use the demo hosted on our landing page at misolabs.ai. To try it locally, follow the instructions below.

If you do not have uv installed yet:

curl -LsSf https://astral.sh/uv/install.sh | sh

Then clone the repository and create the environment:

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

Then run the example conversation. By default, run_misotts.py loads the public model from MisoLabs/MisoTTS and downloads it into the Hugging Face cache if it is not already present on your machine:

uv run python run_misotts.py

The script writes full_conversation.wav in the repository root.

With pip instead of uv:

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

---

Model Introduction

Miso TTS 8B is a text-to-dialogue RVQ Transformer inspired by the Sesame CSM architecture. It generates Mimi audio codes from text and optional audio context, using a large Llama 3.2-style backbone and a smaller autoregressive audio decoder. To find out more about the architecture, read our blog post.

The model is designed for high-quality conversational speech generation. This repository contains the inference code, model definition, and setup instructions for running Miso TTS locally.

Language support: Miso TTS 8B currently supports English only.

---

Model Summary

Item	Value
Model	Miso TTS 8B
Organization	Miso Labs
Task	Text-to-speech
Architecture	RVQ Transformer
Backbone	llama-8B
Audio decoder	llama-300M
Text vocabulary	128,256
Audio vocabulary	2,051
Audio codebooks	32
Audio tokenizer	Mimi
Max sequence length	2,048
Languages	English only

Architecture

Miso TTS 8B uses two transformer components:

• A large backbone transformer that consumes text/audio-frame embeddings.
• A smaller decoder transformer that autoregressively predicts higher-order audio codebooks within each frame.

The backbone accepts interleaved text and audio tokens, allowing it to condition its generations on the conversation history.

---

Usage

Python

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)

Prompted generation

Miso TTS can condition on prior audio for voice cloning. This is optional; the quickstart example above runs without prompt audio.

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,
)

---

Weights

The model weights are hosted publicly on Hugging Face:

uv run python run_misotts.py

The default model repository is MisoLabs/MisoTTS. The first run downloads the model automatically through Hugging Face Hub; later runs reuse the cached copy.

The first run also downloads the SilentCipher watermarking model from sony/silentcipher. If that separate download times out, rerun the command; the Hugging Face cache resumes from files that already completed.

---

System Requirements

Miso TTS 8B is a large model (~8.2B parameters across the backbone, audio decoder, embeddings, and heads). It is not a lightweight CPU model — plan for a high-VRAM GPU for interactive use.

The numbers below are approximate and cover the model weights plus headroom for the Mimi codec, the SilentCipher watermarker, the KV cache, and activations.

Precision	Weights (approx.)	Recommended VRAM	Example GPUs
bfloat16/fp16	~16 GB	24 GB	RTX 3090 / 4090, A5000, L4 (24 GB)
float32	~33 GB	40 GB+	A100 40 GB, A6000 48 GB, H100

CPU: inference runs but is slow. Budget at least ~20 GB RAM for bfloat16 and ~40 GB for float32.

Disk: the first run downloads ~30–40 GB total — the model checkpoint plus the Mimi codec, the SilentCipher watermarker, and the Llama 3.2 tokenizer — into the Hugging Face cache. Make sure you have the free space before starting.

GPU inference defaults to torch.bfloat16. A 24 GB card comfortably fits the bf16 weights; smaller consumer GPUs (4–16 GB) are not sufficient for the full model.

---

Safety

Miso TTS is a speech generation model. Do not use it to impersonate people, create deceptive audio, commit fraud, or generate harmful content.

Generated audio is watermarked by default. If you deploy this model in another application, use your own private watermark key and keep it secret.

---

Links

• Website: misolabs.ai
• Hugging Face: MisoLabs/MisoTTS
• GitHub: MisoLabsAI
• X: @MisoLabsAI