What is whisper_streaming?

Whisper is a speech recognition model from OpenAI that converts spoken audio into text. It works well on pre-recorded files but was not built for real-time use: it waits for a full audio clip before processing. This project, called Whisper Streaming, wraps Whisper to make it work on live audio, printing text as the speaker talks rather than waiting until they finish. The core technique is called local agreement policy. Instead of transcribing in one pass at the end, the system processes small chunks of incoming audio, compares overlapping results to decide which words it is confident about, and outputs those immediately. This produces a rolling transcript with around 3.3 seconds of delay on average, which is low enough for live use cases like conference captioning or live subtitles. It can transcribe in the original language or translate into English at the same time. The software supports several interchangeable processing backends. The recommended one, faster-whisper, runs on NVIDIA GPUs and is the fastest local option. A slower alternative works without special hardware. For people who do not want to run anything locally, there is an option to send audio to OpenAI's API instead, which requires only a small server but incurs per-minute charges. On Apple computers with M1 or M2 chips, a separate backend optimized for Apple hardware is also available. Installation is done via Python package manager commands. The README includes a server mode where the tool listens on a network socket and accepts audio from a connected client, which makes it easier to integrate with other software. Example client scripts are provided for testing. The project notes that as of 2025 it is becoming outdated and points readers to a newer system called SimulStreaming built by the same research group at Charles University in Prague. This was originally published as a research paper in 2023 and demonstrated at a live multilingual conference. The repository is written in Python and the README is in English.