lucasjinreal/Kokoros

SparkTTS Rust speedup inference

Open

#75 opened on Mar 21, 2025

View on GitHub
 (2 comments) (6 reactions) (0 assignees)Rust (105 forks)github user discovery
help wanted

Repository metrics

Stars
 (785 stars)
PR merge metrics
 (PR metrics pending)

Description

Hi, as more and more LLM-based TTS systems pave the way, they show some remarkable results and expressiveness. Our next goal is to support such a model and make it run as fast as llama.cpp.

However, using llama.cpp to support a brand new model is cumbersome. So, let Rust help!

We are planning to support SparkTTS and Orpheus in the future!

We have a drafted inference engine based on Candle. On macOS, we see obvious speedup compared with PyTorch or ONNX. We can run a 0.5B LLM with more than 35 tokens/s on Mac M1. I believe once we push it further, these LLM-based TTS could be even faster.

As such, I strongly recommend that you guys join me if you are interested!

Please fork our current repo for massive development and support!

https://github.com/lucasjinreal/Crane

This project is not officially released yet, but please don't hesitate to join us at any time. SparkTTS and Orpheus are our goals!

updates

2025.03.26: I have supported SNAC audio decoder in Crane, Orpheus should have no obstacles.

Contributor guide