I've been testing Otter.ai for about six months, primarily for transcribing internal engineering syncs and architecture reviews. The main draw was the promise of accurate transcription for meetings filled with cloud-specific terminology.
My verdict: it's decent for general conversation, but struggles significantly with technical jargon. Here's a breakdown from my notes:
* **AWS service names** (e.g., Lambda, ECS, Cost Explorer): These are usually captured correctly, likely because they're common proper nouns.
* **Specific technical terms**: It frequently mangles things like "idempotent," "throughput," "NAT gateway," or "sidecar pattern." You'll get "idem potent," "through put," "nat gateway" (lowercase), and "side car."
* **Acronyms and code**: Handles "API" fine, but stumbles on less common ones. Saying "ECS Fargate" might come out as "E.C.S. Fargate." Any attempt to verbally spell something out or reference a short code snippet is a coin toss.
The speaker diarization and search features are useful for finding discussion points later. However, the accuracy hit on jargon means I spend a fair amount of time correcting key technical sections of the transcript, which negates some of the time-saving benefit.
For non-technical stakeholder meetings, it's a solid tool. For deep technical discussions, the error rate is too high to rely on without manual review. Has anyone else pushed it on highly technical content and found better results with specific settings?
Thanks for this. Your notes on "through put" and "side car" are exactly the kind of detail I was looking for.
I've been wondering if their custom vocabulary feature helps with jargon at all. Have you tried uploading a glossary of your team's terms? I'm curious if it makes a noticeable difference or if it's still hit-or-miss.
You've perfectly captured the core problem with these general-purpose tools. The "through put" example is a classic case of the model not understanding the compound noun. It's treating it like two separate, common words.
I've found the custom vocabulary feature to be a partial solution, but with a significant operational burden. If you're disciplined about maintaining and uploading a CSV of terms, it can improve accuracy for those specific strings. However, it does nothing for sentence context or the pattern you describe. It won't teach the model that "sidecar" is a single concept; it will just force it to transcribe that specific sequence of letters when it hears something close. The moment someone says "side-car" with a slight pause, you're back to square one.
For engineering meetings, I've accepted that a raw Otter.ai transcript is just a first-pass draft. I pipe it into a separate step where a simple script uses a domain-specific dictionary to do a find-and-replace on known jargon. It's an extra hop, but it cuts correction time down considerably.
Extract, transform, trust