# OmniVoice-Triton-API

OmniVoice-Triton-API is a self-hosted ai app. OpenAI-compatible text-to-speech (TTS) API that accelerates the open-source OmniVoice model with Triton kernel fusion and CUDA Graph optimization, delivering 646-language speech synthesis, voice cloning, and voice... It is open source and self-hosted, installable on Unraid.

Canonical page: https://ourlocalhost.com/app/3147/omnivoice-triton-api

## Description

OpenAI-compatible text-to-speech (TTS) API that accelerates the open-source OmniVoice model with Triton kernel fusion and CUDA Graph optimization, delivering 646-language speech synthesis, voice cloning, and voice design at up to 3.4× the speed of stock OmniVoice (~170 ms per utterance) for NVIDIA GPU users.

## Platform availability

Available on 1 of 10 platforms: Unraid.

## Health score

35/100 (Low), based on 4 of 4 factors.
- Maintained: 59 (last commit 164 days ago)
- Popular: 10 (2 GitHub stars)
- Easy to install: 35 (1 install platform, install notes, Docker image)
- Light to run: 10 (8 GB minimum RAM)

## Links

- Unraid: https://ca.unraid.net/apps/omnivoice-triton-api-131plq21vukfqu?q=OmniVoice-Triton-API
- GitHub: https://github.com/hsiang-han/OmniVoice-Triton-API

## Common questions

### Can I point existing OpenAI TTS clients at this service?

Yes, the project exposes an OpenAI-compatible TTS API.

### What step is required before omnivoice-demo commands exist?

Install the package first with pip install omnivoice, or clone the repo and run uv sync. Without that step, commands like omnivoice-demo will not exist on your machine.

### How do I make the OmniVoice demo page reachable from another machine?

Start the demo with omnivoice-demo --ip 0.0.0.0 --port 8001 and keep port 8001 accessible. If port 8001 is not accessible, the web page may not load from another machine.

### Which PyTorch build should I install for Apple M-series Macs or Intel Arc GPUs?

On an Apple M-series Mac, run pip install torch==2.8.0 torchaudio==2.8.0. For Intel Arc GPUs, use the Intel Arc XPU command, because the wrong PyTorch build can make the model run slowly or fail to use the GPU correctly.

### Does the Triton acceleration require extra VRAM?

No, the Triton kernels fuse multiple HBM round-trips into a single pass and reduce memory traffic without any additional VRAM usage.

---
Source: Our Local Host (https://ourlocalhost.com/app/3147/omnivoice-triton-api), the self-hosted and open-source app directory.
