Whisper-Streaming -- real-time mode for Whisper-like models

[Gldkslfmsd]

Whisper-Streaming is a tool that makes Whisper-like models operate in real-time mode. It uses the LocalAgreement streaming policy with self-adaptive latency -- the delay is based on the level of certainty with the output. There's also a demonstration paper with quality-latency evaluation results, published at a peer-reviewed research conference IJCNLP-AACL 2023, and demo video.

The code is open-sourced on GitHub. We received useful feedback from the community, and hope for more open-source collaboration.

(Sorry for announcing it so late after publishing. I didn't notice this discussion.)