Edge0's Audio8 Clones Any Voice Across 23 Languages With Just 600M Parameters
Edge0 released a 0.6B text-to-speech model with zero-shot voice cloning, 23-language support, and competitive accuracy against models up to 8x larger.
Takeaways
- Edge0 released Audio8-TTS-Realistic-Air, a 0.6B TTS model with zero-shot voice cloning, Apache 2.0 licensed.
- Supports 23 languages including Arabic, Hindi, Thai, Vietnamese, adding 12 to the previous Preview release.
- Scores 1.25% error on Seed-TTS, competitive with 4B to 8.5B parameter rivals.
- DualAR architecture: slow transformer picks semantic tokens, fast transformer fills codec codebooks at 44.1 kHz.
- Ships with SGLang Omni adapter exposing an OpenAI-compatible /v1/audio/speech endpoint.
- Quality varies by language; Thai hits 48% WER while German and English stay under 3.3%.
Edge0 has released Audio8-TTS-Realistic-Air, a 601 million-parameter text-to-speech model that supports zero-shot voice cloning across 23 languages. Zero-shot cloning lets the model imitate a speaker from a reference recording and transcript without additional training.
Realistic Air fine-tunes Audio8-TTS-Preview-0.6B, with improved speech realism claimed by Edge0 while retaining the same compact architecture. The Apache 2.0 release bundles the model weights, tokenizer, custom Transformers code, and a neural audio codec that produces 44.1 kHz audio.
Dual autoregression keeps the model compact
Audio8 follows the DualAR architecture used by Fish Audio S2 Pro. A 24-layer autoregressive transformer predicts one semantic token for each audio frame, tracking content and timing across the utterance. A smaller four-layer transformer then predicts the codec codebooks that represent acoustic details such as timbre and intonation.
The codec reconstructs those codebook values into a waveform. Splitting sequence planning from frame-level detail allows the main model to remain relatively small while producing high-resolution audio.
Architecture and runtime specifications
- Main model: 601 million parameters, excluding the codec
- Slow transformer: 24 layers with a hidden width of 896
- Fast transformer: 4 layers
- Codec: 10 codebooks with 4,096 entries each
- Audio: 44.1 kHz output at about 21.5 codec frames per second
- Context: up to 2,048 combined text and audio positions
- Precision: BF16
- Tested hardware: NVIDIA H20
Text, reference audio, and generated audio share the 2,048-position context window, so longer prompts leave less room for synthesis. At 21.5 audio frames per second, 2,048 audio-only positions would represent roughly 95 seconds, although practical output limits are lower once the other inputs are included.