Built-in Domain-Specific Custom Word Packs for Transcription

[jorgedanisc]

I recently saw a demo from AquaVoice here. In the video, they mention achieving better accuracy for domain-specific recordings (for example, software engineering discussions with lots of technical terms).

In this project, there is already a "Custom Words" feature where users can manually add terms they want the model to recognize. My suggestion is to extend this with built-in, toggleable word packs for common domains. For example:

• Software development
• Data science
• Arts
• Construction
• Design
• (and others)

Starting with Software development should be good to validate

Each pack would contain a curated set of relevant terms and could be updated over time as new terminology appears. Users could enable or disable these packs as needed, while still adding their own custom words on top.

This would reduce the need for users to manually build large lists (e.g., hundreds of technical terms) and would help improve accuracy when working with domain-specific recordings.

[cjpais]

Right now the app is not really sophisticated enough for this. If someone wants to implement and bring the feature in that's great and I welcome PRs, but this is a free app!

Post processing was just implemented and the word correction feature built in is not quite good enough imo to support this level of usage

[jorgedanisc]

Sure, I wasn’t suggesting that the app should stop being free — far from it.

It’s just that I played around with “Custom Words” a bit, and it looked promising. That’s why I suggested this feature: instead of manually writing out numerous terms you’d use in software engineering, for example, you could simply have a toggle that, when enabled, would automatically add 500+ terms during the processing step.
That was essentially the core idea.

[cjpais]

Totally, feature creep is just sometimes hard for me to manage as I keep all the issues in my head :) and refactoring too so I can be a little sensitive on this topic

But if you have a good UI/UX in mind I don't see why not.

It's potentially even possible to streamline the whole post processing pipeline

[jorgedanisc]

Sure, I totally get your concerns about feature creep. Anyway, I was just leaving the idea open for discussion because I thought about it when I saw that video. I’ve also been experiencing issues where if I say words like “Tauri” or “TanStack” — you get the idea — the model (Wisper Turbo) I’m using sometimes doesn’t pick up specific terms, even though I find to be an amazing model. So that was a potential solution for that, or at least the beginning of one.
I don’t really have time to work on this at the moment, but if I find time next year, I could try.

Basically, to explain the idea a bit better: we would pre-build a list of predefined words — a vast number, like 500+. For example, software engineering would be one category with nearly a thousand words or so. Then, in the app settings under Custom Words (or maybe above it), you’d have toggles for categories like “Software Engineering.” You could toggle them on or off, and when enabled, those words would be used/added during the processing step.
If this would work, it would be relatively fast to integrate. If not, then maybe it’s not worth pursuing for now because it could get really complex to handle. But I’m just thinking from the top of my head based on the current architecture of the Custom Words feature.