uncloseai.

Qwen3-TTS

Live — This is our default engine, running right now on the public endpoint.

What It Does

This engine clones real human voices from short audio samples. We ship 42+ distinct speakers out of the box — each with their own timbre, cadence, and character, cloned from the LibriSpeech public domain corpus. You can also clone your own voice from a 3-second recording.

It speaks 10 languages natively: English, Chinese, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian. First audio arrives in under 100 milliseconds.

Try It

# Female voice
curl https://speech.ai.unturf.com/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "input": "We rescue abandoned text-to-speech models and give them a new home. No API keys, no tracking, just open source voices for everyone.",
    "voice": "aria"
  }' > aria.mp3

# Male voice
curl https://speech.ai.unturf.com/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Like raccoons digging through digital dumpsters, we find the best open source TTS models that big companies left behind, and we make them accessible to everyone.",
    "voice": "atlas"
  }' > atlas.mp3

from openai import OpenAI

client = OpenAI(
    api_key="not-needed",
    base_url="https://speech.ai.unturf.com/v1"
)

client.audio.speech.create(
    voice="aria",
    input="Raccoon mission. We dig through abandoned repos and rescue the best open source speech models before they disappear."
).stream_to_file("aria.mp3")

client.audio.speech.create(
    voice="atlas",
    input="Five engines, one API. Self-host it, clone any voice, speak ten languages. No vendor lock-in, no API keys, no limits."
).stream_to_file("atlas.mp3")

Voices

42+ built-in voices, each a distinct human speaker. 20 female, 20 male, plus OpenAI-compatible aliases.

Female: aria, clara, elena, grace, hazel, iris, luna, maya, ruby, sage, sofia, amber, brooke, cora, diana, eden, faye, gemma, hope, ivy

Male: atlas, caleb, felix, hugo, jasper, kai, leo, marcus, owen, theo, archer, blake, cole, dane, ezra, finn, grant, heath, ivan, jude

→ Browse all voices live

Technical Details

Model: 1.7 billion parameters
License: Apache 2.0 (commercial use OK)
Languages: 10 (English, Chinese, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian)
Sample rate: 24kHz
Latency: 97ms first packet
Hardware: GPU recommended, runs on CPU (slower)
Upstream: Qwen3-TTS by Alibaba, actively maintained

Self-Hosting

This engine is enabled by default. Clone the repo and deploy:

git clone https://git.unturf.com/engineering/unturf/uncloseai-speech.git
cd uncloseai-speech
make deploy
make voices-qwen   # Download 42+ cloned voice samples from LibriSpeech
make test

The Qwen3-TTS model (~3.4 GB) downloads automatically on first use. The make voices-qwen target runs scripts/download_diverse_voices.py, which pulls speaker samples from the LibriSpeech corpus and assigns them permanent human names via the voice registry.

← Back to Text-to-Speech overview