Audio8-TTS-Preview-0.1B-ONNX-INT8

Audio8 TTS Preview 0.1B — ONNX INT8

Stateful ONNX Runtime export of Audio8/Audio8-TTS-Preview-0.1b, produced for WinSTT. This repository is an unofficial conversion and is not affiliated with Audio8.

The original checkpoint ships no ONNX graphs. Its 24-layer Falcon-H1 slow branch mixes attention and Mamba state, so this export exposes a one-token recurrent contract instead of tracing the complete generation loop:

• slow_step_int8.onnx: semantic logits, slow hidden state, attention KV deltas, and updated convolution/SSM states.
• fast_step_int8.onnx: one within-frame codec-codebook step and attention KV deltas.

Both graphs use per-channel dynamic signed INT8 weight quantization and external data sidecars. The export is pinned to upstream revision 7a644014c398a0495d5efd1da7461bfeb4dbddcd and validated against PyTorch at the first two recurrent positions. WinSTT reuses Audio8's official 0.6B fp16 codec decoder because the 0.1B and 0.6B codec.pth files are byte-identical.

Reproduce

From the WinSTT repository, in an environment containing PyTorch, Transformers 4.57.5, ONNX, ONNX Runtime, and Hugging Face Hub:

python tools/onnx/export_audio8_tts_01.py --output-dir audio8-01-onnx

License

Copyright © 2026 Audio8. The converted weights remain subject to the included Audio8 Community License v1.0. Among other conditions, commercial use is revenue-capped at entities with annual revenue below US$2,000,000, and redistribution must retain the license and Audio8 authorship notice. Review the included LICENSE before use.