Built-in ASR custom vocabulary is real. Here's when you still need a correction layer — and when you don't.
Good question — you should ask it. Every major speech-to-text provider ships some form of custom vocabulary or keyword boosting. If that's solving your problem, you don't need us.
Here's where it stops being enough:
Custom vocabulary features give you a box to type words into — they don't hand you the words. Every new product launch, every regional brand name, every season's additions become your job to track down and keep current. We maintain the agronomy vocabulary as the product. You get the updates, not the upkeep.
Azure's custom vocabulary only helps if you're on Azure. Switch providers, retrain everything. Vocab Layer sits after your ASR step, so it works the same whether your transcript came from Deepgram, Google, Apple's on-device recognition, or anything else. Change your ASR vendor next year — your correction layer doesn't care.
Biasing recognition toward a word list helps, but it doesn't reason about context. When a term is genuinely ambiguous, we run it through a disambiguation pass that looks at the sentence around it — something a static keyword list structurally can't do.
We're not claiming to out-transcribe Deepgram or Google — they're excellent at turning speech into text. We're the layer that makes sure the text is right for an industry neither of them was built to know.