Audio8 TTS 0.1B

Audio8 TTS Preview 0.1B is a small text-to-speech model from Edge0 (Audio8) that supports speech generation and zero-shot voice cloning from a short reference clip. The main generative model has roughly 170M parameters, plus a separate ~120M-parameter codec decoder bundled in the repository. It uses an Audio8 Falcon H1 architecture with a slow autoregressive branch predicting semantic tokens and a fast branch predicting 10 codec codebooks, producing 44.1 kHz audio.

Chinese and English are the primary languages; German, Spanish, French, Italian, Japanese and Korean are listed as experimental. The model card states the aim is practical zero-shot TTS at a small footprint, not identical quality across all languages or benchmarks. Context is limited to 2,048 packed text/audio positions.

Weights are downloadable from Hugging Face and load via Transformers with custom remote code (Python 3.11+, CUDA GPU recommended). A separate ONNX INT8 package supports CPU-only inference with ONNX Runtime, reportedly using about 0.4 GiB after loading, and includes a CLI, web/HTTP service, streaming PCM and voice registration. It is released under the Audio8 Community License v1.0, not a standard OSI license.