nextcloud-talk-capture

Captures Nextcloud Talk calls: the service joins a call, takes one audio stream per speaker, and streams it to a processing gateway.

That is all it does. No recognition, no analysis, no writing files — the finished files come back from the gateway, and the Nextcloud app puts them away under its own credentials, which never leave the customer's premises.

Part of Voxonta. Documentation: running the connector yourself.

Three parts, one of them here

Nextcloud + the app             the connector (this repo)            the cloud
───────────────────             ─────────────────────────            ─────────
keeps the meeting archive       reads its settings from the app
serves settings and calls  ───▶ joins the call
writes the finished files       streams per-speaker audio       ───▶ processing
        ▲                                                            │
        └────────────────── finished files ◀─────────────────────────┘

Nextcloud is a black box to the cloud and the other way round: the connector does not know what happens to the audio once it is sent, and the cloud has no access to Nextcloud.

Install

Two values are needed. Everything else the service asks the app for at startup — the signalling server and its secret, the bot account, which conversations to capture, the folder names.

docker run -d --name nextcloud-talk-capture \
  -e NEXTCLOUD_URL=https://cloud.example.com \
  -e APP_SERVICE_TOKEN=<key from the app's admin settings> \
  -e GATEWAY_TARGET=<gateway address> \
  -e GATEWAY_TOKEN=<gateway key> \
  nextcloud-talk-capture

The service makes outbound connections only: behind NAT it needs no port forwarding and no inbound webhooks.

Where to put it

Capture is a WebRTC client — it connects to the signalling server the way an ordinary participant does. Running it next to the signalling server is not required, but the closer the two are on the network, the fewer reasons ICE has to fail to agree.

Environment

As a library

pip install nextcloud-talk-capture

from talk_capture import AppClient, AppCallMonitor, BrainSink, SpreedClient

The generated contract stubs ship inside the package (talk_capture/_pb), so installing needs no protoc. After editing proto/meeting/v1/meeting.proto regenerate them with scripts/regen_stubs.sh.

Development

pip install -e '.[dev]' pytest
python -m pytest tests -v

Licence

MIT.