Matcha-TTS
Fast TTS with conditional flow matching for efficient speech synthesis.
About
Developed at KTH Royal Institute of Technology and published at ICASSP 2024, Matcha-TTS is a non-autoregressive text-to-speech model that uses optimal-transport conditional flow matching, a technique similar to rectified flows, to cut the number of ODE solver steps needed for synthesis. The result is fast, probabilistic speech generation with a compact memory footprint and natural-sounding output, with controls for speaking rate, sampling temperature, and solver step count to trade speed against quality. The package installs with pip and offers a command line tool, a Gradio web demo, and a Jupyter notebook, plus full training scripts for custom datasets and ONNX export for deployment. Released as open source under the MIT license, it serves speech researchers as a flow-matching baseline and developers who need lightweight local TTS, and its approach has influenced later speech systems that adopt flow matching.
Should you use Matcha-TTS?
Pick it when
Pick Matcha-TTS when you are training your own voice on a custom dataset and want a fast MIT licensed flow-matching baseline with ONNX export, speaking rate and solver step controls, and a small memory footprint.
Look elsewhere when
Skip it if you want ready-made voice cloning or many languages without training; F5-TTS or Chatterbox clone zero-shot. Its research-oriented setup also means more work than a pip-and-go model like Kokoro.
Alternatives to Matcha-TTS
- VITS
Widely forked MIT end-to-end baseline with a stochastic duration predictor, but older, with Cython build steps and no flow-matching speedups.
- StyleTTS 2
Rated above human recordings on LJSpeech via style diffusion and adversarial training, a higher quality ceiling, but its models carry a synthetic speech disclosure rule.
- F5-TTS
Flow matching scaled up to zero-shot cloning from a short clip, usable without training, but its pretrained checkpoints are non-commercial.
- Piper TTS
Trains custom VITS-based voices and deploys them on a Raspberry Pi with an HTTP server, the practical route when you want a shipped voice.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- Text-to-Speech (TTS)
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Advanced (4/5)
- License
- MIT
- Minimum VRAM
- 4 GB
- Added
- Apr 3, 2026