GitHub radar
Audio8 ASR Infinite: an infinite stream of speech recognition models
Edge0 has published a long-term stable streaming ASR model which prevents the long audio from being messy at the end of the recording.
A long-term stable streaming ASR model that avoids messy endings for long audio recordings. Audio8 ASR Infinite is an online speech recognition model created by Edge0 that can convert audio signals into corresponding texts in real time. By using the streaming model, it allows us to achieve a high decoding frequency (12.5Hz). It also allows us to set different delays (240ms–560ms) and recognizes both Chinese and English languages. This model is quite small, only containing around 4 billion parameters. But unfortunately, because this model is not a general-purpose model, we can only deploy it locally using Docker to use it as a light-weight streaming service. We also have a demo website available!
Why a vibe-coder should care
It is intended for use with long audio such as interviews, live streams, meetings, etc. so that there won’t be any problems caused by long audio during transcription. You can get near real-time subtitles with low latency (<1s), instead of losing the rest of your speech. But unfortunately, this model requires a Docker environment instead of just installing an application.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the Audio8 ASR Infinite model from Edge0 for me: https://huggingface.co/Edge0/Audio8-ASR-Infinite — follow their Docker setup, launch the web demo, and show me how it transcribes my speech in real time.
A regular laptop with 8 GB of RAM or more can hold the model — the weights take about 8 GB. For real-time transcription a GPU helps a lot: the developer's own examples run it on one, and on a CPU alone the response lags noticeably.
Open on Hugging Face▌ More finds