3v324v23's picture
Disable incremental waveform streaming
835ac16
|
Raw
History Blame Contribute Delete
1.41 kB
metadata
title: Audio8 TTS Preview 0.6B
colorFrom: red
colorTo: gray
sdk: docker
app_port: 7860
models:
  - Audio8/Audio8-TTS-Preview-0.6b
license: apache-2.0

Audio8 TTS Preview 0.6B

Multilingual text-to-speech and zero-shot voice cloning for Audio8/Audio8-TTS-Preview-0.6b. The service uses the official Audio8 adapter with SGLang Omni 0.1.0 and SGLang 0.5.8.

Reference audio uploaded through the web interface is limited to 30 seconds. The A10G runtime enables torch.compile with SGLang's native batch policy and uses CUDA Graph without an Audio8-specific compile batch-size limit. Waveform decoding runs once after acoustic-token generation; incremental waveform streaming is disabled for long-form stability.

API

curl -sS \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "audio8/tts-0.6b",
    "input": "Audio8 generates this sentence with the reference voice.",
    "response_format": "wav",
    "temperature": 0.8,
    "top_p": 0.95,
    "top_k": 50,
    "references": [{
      "audio_path": "https://example.com/reference.wav",
      "text": "The exact transcript spoken in the reference audio."
    }]
  }' \
  https://audio8-audio8-tts-preview-0-6b.hf.space/v1/audio/speech \
  -o audio8-clone.wav

Use reference voices only with the speaker's permission and disclose synthetic audio where appropriate.