--- title: duplex-0 emoji: 🎙️ colorFrom: blue colorTo: gray sdk: gradio sdk_version: 6.32.0 app_file: app.py short_description: Small full-duplex speech-to-speech model (Mimi tokens) python_version: "3.12" startup_duration_timeout: 30m models: - Trelis/duplex-0 - kyutai/mimi - kyutai/moshiko-pytorch-bf16 --- Demo of [Trelis/duplex-0](https://huggingface.co/Trelis/duplex-0), a 201M-parameter two-track speech-to-speech research model. The model code in `duplex/` is copied from the model repo's `scripts/duplex/model` (Apache-2.0, see `duplex/NOTICE`). Example prompts are the first 10 seconds of the "real" sample clips in the model repo's `report/samples/` (Apache-2.0 repository). Audio is encoded/decoded with Kyutai's [Mimi](https://huggingface.co/kyutai/mimi) codec, and text uses Moshi's tokenizer.