# Matcha-TTS: A fast TTS architecture with conditional flow matching

**Authors:** Shivam Mehta, Ruibo Tu, Jonas Beskow, Éva Székely, Gustav Eje Henter  
**arXiv:** [2309.03199](https://arxiv.org/abs/2309.03199)  
**Date:** 2023-09-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.03199v1/

## Key points

- Matcha-TTS is a novel encoder-decoder architecture for Text-to-Speech (TTS) acoustic modeling.
- It focuses on speed and efficiency in TTS synthesis.
- Utilizes optimal-transport conditional flow matching (OT-CFM) during training.
- ODE-based decoder produces high-quality output with fewer synthesis steps compared to score matching models.
- Each synthesis step is fast to execute, making it ideal for real-time applications.
- Probabilistic and non-autoregressive, can generate speech without external alignments.
- Outperforms strong pre-trained baseline models in several aspects:
- Smallest memory footprint, efficient resource usage.
- Rivals the speed of fastest models for long utterances.
- Achieves highest mean opinion score in listening test, indicating superior audio quality.
- Audio examples, code, and pre-trained models available for exploration at https://shivammehta25.github.io/Matcha TTS/.
- Comprehensive resource for researchers and developers to integrate Matcha TTS into their projects.
- Unique approach using OT CFM sets it apart from traditional methods.

## Summary

Matcha-TTS is a novel encoder-decoder architecture for Text-to-Speech (TTS) acoustic modeling that focuses on speed and efficiency. It utilizes optimal-transport conditional flow matching (OT-CFM) during training, resulting in an ODE-based decoder that produces high-quality output with fewer synthesis steps compared to models trained using score matching. One of the key advantages of Matcha-TTS is its careful design choices which ensure that each synthesis step is fast to execute, making it an ideal choice for real-time applications where speed is crucial. Additionally, Matcha-TTS is probabilistic and non-autoregressive meaning it can generate speech without relying on external alignments. In terms of performance, Matcha-TTS outperforms strong pre-trained baseline models in several aspects; it has the smallest memory footprint making it more efficient in terms of resource usage; it rivals the speed of the fastest models when processing long utterances; and achieves the highest mean opinion score in a listening test indicating its superior audio quality. For those interested in exploring Matcha TTS further, audio examples, code and pre trained models are available at https://shivammehta25.github.io/Matcha TTS/. This comprehensive resource allows researchers and developers to experience the capabilities of Matcha TTS firsthand and integrate it into their own projects. Overall, Matcha TTS presents a compelling solution for TTS acoustic modeling by combining speed, efficiency and high quality output. Its unique approach using OT CFM sets it apart from traditional methods and opens up new possibilities for real time speech synthesis applications.