OPEN-OMNIVOICE-TTS

A sentence in 0.3 seconds. Any voice, 600+ languages, on your own GPU.

A self-hosted text-to-speech server with an OpenAI-compatible API, built on OmniVoice, k2-fsa's zero-shot voice cloning model for 600+ languages. It runs in Docker on an NVIDIA GPU, and any OpenAI client can use it: Open WebUI, LibreChat, SillyTavern, AnythingLLM, or the openai SDKs.

• OpenAI-compatible /v1/audio/speech: all output formats, speed, and all 13 OpenAI voice names
• Voice cloning from a 3 to 20 second recording, and 12 open voices ready to use
• OpenAI's custom voice API: create voices from an audio sample or from a description
• Fast: 25 to 30 times faster than real time on an RTX 3080, about 3.4 GB of VRAM
• Interactive API docs at /docs

Examples

GitHub cannot play audio in a README: click a file to download it, or listen to all of them on the project page.

Cloning. Listen to the original, then the clone. The clone says words that were never recorded; everything else about the voice comes from the 15-second original.

The bundled voices. Twelve of OpenAI's voice names come with an openly licensed voice, cloned from LibriVox readers in LibriTTS-R (credits: Voices/ATTRIBUTION.md):

Requirements

Check that Docker can see your GPU:

docker run --rm --gpus all nvidia/cuda:12.8.2-base-ubuntu24.04 nvidia-smi

Quick start

git clone --recursive https://github.com/hypersniper05/open-omnivoice-tts.git
cd open-omnivoice-tts
./start.sh            # Linux

.\start.cmd           # Windows (or double-click start.cmd)

The first start builds the Docker image for your GPU (it downloads about 15 GB and takes 10 to 30 minutes), then downloads the model (about 3 GB). When the server is ready, the script prints:

Ready after 95 s. OpenAI-compatible API (no API key):
    base URL   http://localhost:8008/v1
    API docs   http://localhost:8008/docs   (try POST /v1/audio/speech there)

Try it:

curl http://localhost:8008/v1/audio/speech -H "Content-Type: application/json" \
  -d '{"model": "tts-1", "voice": "alloy", "input": "Hello from open-omnivoice-tts."}' --output hello.mp3

• Stop: ./stop.sh or stop.cmd. Logs: docker compose logs -f.
• More than one GPU: before the first start, copy .env.example to .env and set OMNIVOICE_GPU to the nvidia-smi number (or UUID) of the card to use. The default is GPU 0.
• Updates: after pulling new code, run ./start.sh --build or start.cmd -Build.

Security note. By default there is no API key, and the server listens on all network interfaces. Set api_key in app/openai_voices.json before other computers can reach port 8008, use a VPN such as Tailscale for remote access, and never forward the port to the internet.

Connect a client

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8008/v1", api_key="not-needed")
client.audio.speech.create(model="tts-1", voice="nova", input="Hello.").write_to_file("hello.mp3")

Open WebUI: Admin Panel > Settings > Audio: engine OpenAI, base URL http://<host>:8008/v1, any key, model tts-1. Other clients: choose OpenAI as the text-to-speech provider and set the same base URL.

API

• voice: a voice name, a file in Voices/, a custom voice id, or design:<attributes> (e.g. design:female, young adult, british accent). Also accepted as {"id": "..."}.
• input: up to 20,000 characters. Tags such as [laughter] and [sigh] work in the text.
• response_format: mp3 (default), opus, aac, flac, wav, pcm. speed: 0.25 to 4.0.
• Errors use OpenAI's format. A busy server answers 429 or 503 with Retry-After. Not supported: stream_format: "sse".

Voices

Put a clip of one speaker (3 to 20 s, little background noise) in Voices/ and use its file name as the voice.

• Transcript: put what the clip says in <name>.txt next to it. The clone is better, and the server does not need to run Whisper.
• Silence: add 0.5 s of silence before the speech and 0.3 s after it; this prevents a breath or blip at the start of the speech: ffmpeg -i clip.wav -af "adelay=500:all=1,apad=pad_dur=0.3" -ar 24000 -ac 1 clip_padded.wav
• Names: app/openai_voices.json maps voice names to files. Edits apply without a restart.
• More open voices: tools/pick_libritts_voices.py picks clean clips from a LibriTTS-R download and writes their credits.

Custom voices (OpenAI's voice API) are off by default: set "allow_custom_voices": true and an api_key in app/openai_voices.json. A voice from a description uses OmniVoice's attributes (gender, age, pitch, style, accent):

voice = client.audio.voices.create(type="prompt", name="Narrator",
                                   prompt="female, middle-aged, low pitch, british accent")
client.audio.speech.create(model="tts-1", voice={"id": voice.id}, input="Hello.").write_to_file("hello.mp3")

A voice from an audio sample needs a consent recording first (POST /v1/audio/voice_consents); see /docs.

Only clone voices you have the right to use.

Configuration

app/openai_voices.json (API settings, no restart needed):

.env (server settings; run the start script again after a change):

Machine-specific Docker settings, such as CPU or RAM limits, go in docker-compose.override.yml (git ignores it).

Performance

RTX 3080, default 16 steps. A sentence is about 4.6 s of speech, a paragraph about 32 s:

CPU mode (OMNIVOICE_DEVICE=cpu) works but is more than 100 times slower. Whisper, which transcribes voices that have no transcript, had a 1.7% word error rate on the bundled clips.

Troubleshooting

• No GPU found: run the nvidia-smi check above. On Linux, install the NVIDIA Container Toolkit.
• MissingJITCacheError: the image was built for another GPU. Run start.cmd -Build or ./start.sh --build. Only RTX 30-series cards have been tested.
• A breath or blip at the start of the speech: add silence to the voice's clip (see Voices), or use a clip with no breath before the first word.
• Other computers cannot connect: use this computer's IP address and allow port 8008 in the firewall.
• Harmless: pip's dependency resolver error during the build, and the "unauthenticated requests" warning while the model downloads.
• Anything else: docker compose logs --tail 100.

Credits and licenses

The code in this repository is MIT licensed (see LICENSE). Other parts keep their own licenses:

The OmniVoice model allows non-commercial use only: read its license before you use speech made with it commercially.