Repository navigation
Highlights
Bounded audio backlog per session (#64, #65)
While transcription was busy, every incoming WebSocket audio frame was queued without limit. WF_MAX_WINDOW_CHUNKS only applied after audio left the queue, so a client sending faster than real time, or a slow model, could grow server memory without bound.
Each session now refuses audio once its queue holds WF_MAX_WINDOW_CHUNKS frames or WF_MAX_QUEUED_BYTES bytes, and the server closes the socket with the new close code 4001 (audio backlog full). Audio still queued at that point is discarded.
WF_MAX_QUEUED_BYTES: new, default8388608(8 MiB, about 4 minutes of 16 kHz int16 audio)- Flush and stop markers bypass the limit, so they still go through when the queue is full
This is an additive protocol change. PROTOCOL_VERSION stays at 1. See docs/protocol.md.
Full changelog: v1.3.0...v1.4.0
Install
pip install -U whisperflow==1.4.0