AI product Open source
Speakrail is a fully local, full-duplex voice assistant that runs on a single 24 GB NVIDIA GPU such as an RTX 4090. It combines Voxtral Mini 4B Realtime for speech recognition with a turn-taking head that evaluates the conversation every 80 ms, and uses Gemma 4 12B with a Speakrail LoRA to decide whether to speak, listen, interrupt, yield, or interject. Breeze TTS 2 produces the assistant's speech; peek decoding can begin preparing a response before the turn is fully confirmed, while listening notes preserve longer or corrected requests. The local web application provides tools for weather, timers, notes, lists, calculations, unit conversion, and optional web search or external model access. It is distributed as Docker Compose services and requires Linux, Docker with NVIDIA Container Toolkit, roughly 75 GB of disk space, and a suitable NVIDIA GPU. Speakrail's code is licensed under Apache 2.0, while the default Breeze TTS 2 model and generated audio are restricted to research and non-commercial use.
1 use taken from transcripts — each links to the moment in the video.
A fully local voice assistant that listens while it talks, running on one RTX 4090. You can cut it off mid-sentence, backtrack, or pause to think, with a small turn-taking model deciding every 80ms whether you're done or just back-channeling.
1 in the library.