Speechcraft

The problem

Speechcraft grew out of services at a local Indonesian community church in Hobart. Attendees who did not speak Indonesian often struggled to follow the sermon, while professional interpreters and dedicated receiver hardware were beyond what a small community could reasonably maintain.

The project works with equipment already in the room. A speaker uses a phone as the microphone and capture console, and audience members open a browser page on their own phones. They do not need to install an application or create an account.

How it works

Audio from the speaker’s browser streams to Deepgram for Indonesian speech recognition. A Supabase Edge Function then translates each segment into English, using recent context and a small glossary to handle pronouns, names, and church-specific language more consistently. Supabase Realtime broadcasts the result to connected audience devices.

The caption stream is ephemeral. Speechcraft does not write transcripts to a database after the service, which keeps the architecture small and reduces the amount of spoken material retained by the product.

Translation quality

Literal phrase-by-phrase translation often produces awkward English because Indonesian sentence structure, implied subjects, and religious terminology do not map cleanly one segment at a time. The current work focuses on preserving enough recent context to make the English readable without delaying the live stream unnecessarily.

The public repository includes the speaker console, audience view, translation pipeline, realtime delivery code, interface screenshots, and an architecture diagram.

Current status

Speechcraft is an open prototype used in the setting that motivated it, but it is still being tuned. Recognition errors, natural grammar, context length, and domain terminology remain active parts of the work. It is a practical captioning experiment, not a replacement for a professional interpreter.