Status & Features

[lucasjinreal]

Hello, all those interested in Kokoro!

Kokoro is gaining popularity in text-to-speech (TTS) due to its small size and extremely high quality. However, making Kokoro run more easily, especially with languages like Rust and C++, is still an area that has not been fully explored. With the help of buildable languages, Kokoro can run on various platforms such as embedded systems, phones, and also on PCs and laptops without invoking any scripts. As a result, “Kokoros” emerges.

Here are the goals of Kokoros:

• Achieve high-quality TTS on embedded systems, PCs, laptops, or any cross-platforms.
• Provide TTS for multiple languages. In the long run, Kokoros will support more variants of TTS, not just Kokoro.
• Implement local large language model (LLM) + TTS.

The current implemented features and plans are as follows:

• Fully functional on Rust, including tokenizer, phonemizer, and styles.
• More comprehensive feature examples.
• Full support for languages such as Chinese, Japanese, German, etc. (English is already supported).
• API support in Rust, similar to the OpenAI API, so that users can run “koko server” to start a TTS OpenAI server.
• Support more variants, not just Kokoro, but any TTS model built upon StyleTTS-V2.
• Docker container support, anyone want contribute for this?

Leave your desired features or contributions below if you are interested.

Some advanced feature that currently doing:

• AMSR style mixing, wants a comfortable voice? How voice blending can be implemented? #10 tracking this

[satvikpendem]

Can it run on the GPU via CUDA or is it only CPU inferencing for now? I believe Rust has CUDA bindings via crates as well, might be a possibility in the future if not now.

[altunenes]

Can it run on the GPU via CUDA or is it only CPU inferencing for now? I believe Rust has CUDA bindings via crates as well, might be a possibility in the future if not now.

Never tried with NVIDIA but I think it should work with Cuda?

Kokoros/Cargo.toml

Line 12 in 1778be2

ort = { version = "2.0.0-rc.4", features = ["coreml"] }

then adjust the EP to work with CUDA

https://github.com/lucasjinreal/Kokoros/blob/main/src/onn/ort_base.rs

[lucasjinreal]

Yes. CUDA is possible, even though CPU is fast enough. Would you guys interested to add a PR to support CUDA ep provider as alternative to device target?

[lucasjinreal]

video-1737110239209.webm

[Jerboas86]

Docker container support, anyone want contribute for this?

I can push a PR on that

[lucasjinreal]

@Jerboas86 that would be awesome, we really need a single command make users can run in docker and with minimal docker container size!