About Terpside

Kevin · interpreter

I am a Spanish ↔ English medical interpreter. I work from Bolivia, over the phone and by video, for clinics and hospitals in the United States through over-the-phone interpreting agencies. I interpret appointments, results, discharge instructions, pharmacy calls: conversations where a misheard number matters.

For a long time my second monitor was a stack of three tools that were not built for this: Windows captions, Chrome captions and a translator on top. They rewrote themselves while I read, showed only two or three sentences, and if I missed a number, it was gone.

In July 2026 I started building what I wanted to have: a single window that waits for the person to finish talking, shows the turn in its language with the translation underneath, and does not save the text, because nothing from a clinical call should stay on a disk. I have used it every shift since, and I measure it against my own calls.

That is Terpside: an interpreter's tool, shared with other interpreters. I sell it because colleagues asked me about it; the Free plan exists so anyone can try it on real calls without paying, and Pro is for whoever wants it all shift long.

— Kevin Cerpa, medical interpreter and creator of Terpside
Jul 30, 2026
in daily use since
2,000–3,000
turns per shift
3 shifts
audited: every published turn was archived (Sep 16–18)
0
internet connections

Numbers from my own computer (Ryzen 5 3600, 32 GB, GPU) and my own calls, July to September 2026. They are my numbers, not a guarantee for your machine; the public installer is measured before it is published. The audit of archived turns used a separate tool, authorized on my own computer; the edition that is sold stores nothing.

How I build it

Terpside is written in Python and relies on open models that run on your computer: Whisper (OpenAI, MIT license) to recognize and Opus-MT (University of Helsinki, CC BY 4.0) to translate. I chose Opus-MT over the alternative I tested (NLLB-200) because in my tests it translated the trade's vocabulary better, used one eighth of the CPU resources and its license allows selling it; full attributions are in the "About" screen.

I measure each change before keeping it: latency, CPU use, missing numbers and altered negations. When a test shows an idea does not work, it is dropped even if it sounded good.

Questions? Write to me

I reply myself. If you work differently than I do, I would like to hear about it.

Support and contact