docs: STT research — verified Deepgram + whisper.cpp findings and decisions
Two researcher passes (2026-09-17), cross-checked against our codebase:
Deepgram streaming (WebSocket, Nova-3, SDK v5, $0.0048/min streaming
promo, $200 credit), AudioWorklet linear16 capture (MediaRecorder webm
has unresolved empty-transcript reports), endpointing/utterance_end/keep-
alive gotchas, 127.0.0.1 secure context, server-minted short-lived access
tokens (renderer never holds the real key); whisper.cpp CLI child process
(Metal RTF base 0.04/small 0.08, base.en 142MB/small.en 466MB, MIT) with
per-utterance batch flow and LH_DATA_DIR model cache. Decisions: toggle
mic, cloud-first wave scope, renderer->wss direct.
amosblomqvist
6e08d77ec2