nextcloud-talk-capture 0.2.0

Capture Nextcloud Talk calls: signalling client, per-speaker audio, call monitoring

nextcloud-talk-capture

Capture Nextcloud Talk calls: join as a signalling client, receive one audio stream per speaker, and hand over the finished call as plain data.

Import name is talk_capture.

from talk_capture import CaptureConfig
from talk_capture.monitor import CallMonitor
from talk_capture.spreed_client import SpreedClient

What it does — and what it deliberately does not

This is the capture half of a transcription pipeline: it gets the audio and tells you who was speaking. What happens afterwards — speech recognition, summaries, publishing, storage — is not its business, and it holds no dependency on any of it.

That separation is the whole point. The same package runs inside a Nextcloud ExApp, inside a standalone capture service, or inside an application that does everything, without dragging a product's worth of code along.

Module
spreed_client	signalling client: HPB + aiortc, per-speaker audio
monitor	active-call detection (polls the Nextcloud DB, optionally over an SSH tunnel)
presence	join/leave annotations (pure functions, no I/O)
chat_commands	opt-out commands posted in the room chat, honoured mid-call
resolve	assembles a FinishedCall when the call ends
transcriber_backend	sends audio to an engine, in-process or over gRPC
contracts	FinishedCall, SpeakerStatInput — the payload handed downstream
config	CaptureConfig — what capture needs, nothing more
aioice_patch	works around an aioice race; import it before aiortc

Requirements

• Nextcloud Talk with a High Performance Backend. Call media lives in the signalling server, not in the PHP layer, so an HPB is required.
• Python 3.11+

Configuration

CaptureConfig carries what capture needs: Nextcloud credentials, the signalling server and its secret, call-detection database access, the room filter, and where to send the audio. Build it from the environment with CaptureConfig.from_env(), or pass any object exposing the same attributes — config is taken duck-typed, so a larger application config works unchanged.

Contributing

Two rules keep the package shippable on its own, both enforced by tests in the consuming repository:

• nothing here may import the downstream half (publishing, analysis, storage);
• a config field must keep its name, type and default in sync with any larger config that is passed in — a mismatch fails mid-call, not at import.

Licence

MIT.