Radio call-ins in Ugandan languages, as themes on a dashboard

Pipeline design · 2026

stages, each with a swappable model

6

stages, each with a swappable model

per new model in the benchmark

1 column

per new model in the benchmark

against the anthropologists' gold set

WER

against the anthropologists' gold set

TRAC FM runs radio polls in Uganda. Listeners call in and answer in their own languages, often over noisy lines. I designed the pipeline that turns those clips into English text, groups the answers by meaning, and shows them per question and region.

At a glance

• Model-agnostic by design: cleaning, transcription, translation and grouping are each swappable.
• A benchmark harness: add a provider and a model, and it runs the gold set and appends a column, with WER for transcripts and expert-validated translations as the reference.
• Containerised services from ingest to dashboard, so a better model is a configuration change.