Whisper Streaming is a real-time speech-to-text transcription and translation system that extends OpenAI's Whisper model for streaming audio processing.
The tool addresses Whisper's limitation of being designed for batch processing rather than live transcription. It implements a local agreement policy with self-adaptive latency to enable streaming transcription of long-form, unsegmented audio. This approach allows the system to output transcriptions incrementally as audio arrives, rather than waiting for complete audio segments.
The project suits developers building live transcription services or applications requiring real-time speech recognition and translation across multiple languages. It offers flexibility in backend selection: faster-whisper provides GPU acceleration for performance-critical deployments, whisper-timestamped offers a less restrictive alternative with slower processing, and the OpenAI API backend eliminates local GPU requirements at the cost of per-request fees. The README notes that the project is being superseded by SimulStreaming, which developers should evaluate as an alternative for new projects.
The tool has been demonstrated in production at a multilingual conference and achieves approximately 3.3 seconds latency on long-form speech transcription. Development appears to have stabilized following the publication of the demonstration paper, with the maintainers now directing users toward the successor project for ongoing development and improvements.
