WhisperLive is an open-source project from Collabora that turns OpenAI’s Whisper speech recognition model into a nearly-live transcription system. Published on GitHub, it has attracted about 4,327 stars and was still actively updated in 2026. Where the original Whisper works on complete audio files, WhisperLive streams audio in chunks and produces transcripts with low latency, making it practical for live captions, meeting transcription, dictation, and broadcast tooling such as OBS.
The project is built around a client-server design: audio clients capture microphone or media input and stream it to a WhisperLive server, which runs the Whisper model on accelerated hardware and returns text as it is recognized. Multiple inference backends are supported so the server can take advantage of whatever silicon is available, from NVIDIA GPUs with TensorRT to Intel hardware with OpenVINO and AMD GPUs with ROCm. That flexibility makes it usable both on a workstation and on dedicated transcription servers. Because transcription happens on infrastructure you control, sensitive audio never has to leave your network, which matters for healthcare, legal, and internal corporate recordings.
WhisperLive is completely free and open source under the MIT license. There are no subscriptions or per-minute transcription fees; your only cost is the hardware you run the Whisper models on.
WhisperLive is for developers, streamers, and IT teams who need live speech-to-text they fully control — for example captioning live events, transcribing meetings on private infrastructure, or adding voice recognition to products without sending audio to a third-party cloud service.