Infrastructure¶

The supporting pieces: the media gateway, the relay, and the processing chain that turns a recorded interview into an analysable dataset.

Janus-Gateway¶

A pinned, reproducible container build of the Janus WebRTC Server with layered configuration. Repository · Documentation

Janus is the SFU that carries audio and video between the headset and the browser, plus the data channel used for control messages. FACES maintains its own image for three reasons:


Pinned	Janus and every source dependency are fixed to exact versions. Images that track `master` build a different gateway every time, which is untenable when you have to state which software produced a dataset.
Layered configuration	Stock Janus configs ship inside the image; you mount only the files you change, instead of vendoring 40 KB of defaults into every project.
Complete	Data channels (`usrsctp`) and recording post-processing (`janus-pp-rec`) are compiled in. Both are easy to leave out and painful to discover missing.

docker run -d --network host \
    -v "$PWD/conf.d:/etc/janus.d:ro" \
    ghcr.io/texttechnologylab/janus-gateway:1.4.2

Use host networking

Not required, but strongly recommended. ICE must see the real interface to gather usable candidates; bridge networking gives you a gateway that negotiates successfully and then carries no media.

coturn¶

An optional self-hosted TURN relay. Not needed on permissive networks, but frequently decisive on the university and institutional networks where this research actually runs, behind NATs and firewalls, where a direct connection often cannot be established at all. Configuration lives in the InterView repository; see the coturn guide.

InterView-P and the processing chain¶

InterView-P is a small set of standalone processes that run after an interview, not during it:

Collection. Pulls together everything recorded for a given token across the media server and the database.
Speech recognition. Transcription via Whisper or Whisper-derived models, producing the word-level records that link to audio chunks in the data model.
Gaze resolution. Replays captured eye tracking inside the Unity scene and raycasts it against the actual colliders, producing dwell and fixation measures tied to named areas of interest rather than to screen pixels.
Questionnaire import. Pulls responses from the survey tool (LimeSurvey in production) and synchronises them onto the interview timeline, so answers and behaviour share a clock.

The resulting relational dataset can be handed to DUUI, the Docker Unified UIMA Interface, for distributed processing across a large library of extensible NLP and multimodal tools.

Reproducibility¶

The analysis package that produced every number and figure in Studies regenerates from cache in one command, and its summary document is written by reading the results back out of the generated tables, so the prose cannot drift from the data. Only the cache-building step touches the network; everything downstream is offline and deterministic.

That property is deliberate and, for an infrastructure programme, is part of the deliverable rather than a nicety: a claim about an interview corpus is only as good as the ability to rebuild it.