MisoLabs' MisoTTS Brings Conversational Speech Directly To Your Machine
MisoTTS is a new 8 billion parameter text-to-speech model now available on Hugging Face. It converts written text into natural, conversational speech while maintaining voice consistency from short audio samples. The model uses a Sesame CSM architecture that splits the workload between two transformer components for high-quality output.
Miso Labs developed this model to push forward the state of local speech generation. They combined a large Llama 3.2-style backbone with a smaller autoregressive audio decoder. The result is a system that handles both speech synthesis and voice continuation from prompt audio.
Conversational speech for local machines
Key Features
- Generates speech from text and audio context.
- Uses 32 audio codebooks through Mimi tokenizer.
- Split transformer design for efficiency and quality.
- Handles sequences up to 2,048 token length.
- Supports voice continuation from short audio prompts.
- Runs fully on local hardware configurations.
Privacy-focused professionals and small creative agencies stand to benefit most from this release. Running speech generation locally means sensitive scripts and voice data never leave your machine. Serious hobbyists with capable GPUs can also experiment with high-quality conversational voices without subscription fees or cloud dependencies.
What to know before running
The model is available through inference code on the Miso Labs website and a live demo at misolabs.ai. The backbone handles text and audio embeddings while a smaller decoder predicts higher-order audio codebooks within each frame. Setup instructions and model definitions are included in the Hugging Face repository for local deployment.
"Miso TTS 8B is a text-to-speech model based on the Sesame CSM architecture, generating Mimi audio codes from text and optional audio context, using a large Llama 3.2-style backbone and a smaller autoregressive audio decoder." — Source: Hugging Face
Project Links
