Faster transcription, Deepgram learned vocabulary, Electron 39
Faster on short clips. We now skip the AI cleanup pass on very short
utterances where it adds latency without changing the output.
One less resample. A redundant resampling step in the audio path is
gone — about 340ms shaved off the Deepgram path.
Deepgram learned vocabulary. When you teach Shruvi a custom word, we
now pass it as a biasing keyterm to Deepgram automatically. (Other
providers were already supported.)
Electron 39 runtime. Bumped from Electron 30 to 39 to track the
active LTS support window.
Connection-per-IP limit. Backend now caps WebSocket connections per
source IP via SHRUVI_WS_CONNECT_PER_IP_LIMIT.