What it does
Tortoise turns text into speech and was designed around two priorities: handling many different voices and producing realistic prosody and intonation. The repository holds the code needed to run it in inference mode, and the model weights are hosted on Hugging Face. It combines an autoregressive decoder with a diffusion decoder. The author describes the original design as very slow, though the README now reports around 0.25 to 0.3 real-time factor on 4 GB of VRAM and under 500 ms latency with streaming. You can install it with pip, a conda environment or Docker, and use it through scripts such as do_tts.py and read.py, a Python API, or a socket streaming server. Running it locally requires an NVIDIA GPU, and the README has notes for Windows and for Apple Silicon Macs. A demo on Hugging Face Spaces is linked, and the README also says the project no longer seems to be active.
Capabilities
- Speaks text in multiple voices, with voice samples supplied as reference clips
- Autoregressive and diffusion decoders for realistic prosody and intonation
- Command-line scripts and a Python API
- Socket streaming server for lower latency
- Speed presets plus DeepSpeed, KV cache and half-precision options
- Installs with pip, conda or Docker
When to use it
- You want an open-source text-to-speech model you can run on your own NVIDIA GPU
- You need to read long text files aloud and re-generate clips that came out badly
- You are experimenting with speech synthesis from Python code
