qwen_asr has no quantized-weights option, so the only lever available to
reduce the 1.7B model's memory footprint is avoiding concurrent instances.
Previously each track (system, mic) ran its own transcribeTrack goroutine
with independent qwen_asr subprocesses, so simultaneous speech on both
tracks meant two ~6.9GiB model instances alive at once - over half the
24GB target machine's memory. Merge both tracks onto a single worker
(asrSegs channel, dispatch() routes both tracks onto it) so invocations
are strictly serialized; per-track order is preserved since VAD's output
is already chronological. Verified via pgrep -x qwen_asr polling during a
real dual-track capture: never more than 1 concurrent process.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Go orchestrator: spawns audiotee for capture (system audio + mic, two
tracks), segments with an energy-threshold VAD, transcribes each segment
via qwen-asr (subprocess per segment, vendored as a pinned git submodule
in third_party/qwen-asr), and serves the live transcript over SSE while
also writing it durably to a text file. Includes a glossary/prompt-leakage
guard (internal/asr/leak.go) that discards a segment if the model echoes
the biasing prompt instead of transcribing.
cmd/transcriptor-ai is the main orchestrator; cmd/transcript-tail is a
minimal terminal SSE client. See CLAUDE.md for the full architecture and
the reasoning behind each choice (Go over Python/Swift/Rust, SSE over
WebSocket, why the VAD is a hand-rolled heuristic, etc).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
sttlab-tech