Been using MeetGeek for a few weeks to record my team's infrastructure planning calls. The transcription part is actually pretty solid—I can find when we talked about that broken S3 policy.
But the automatic tagging? It feels completely useless for technical discussions.
It tags things like "project", "deadline", "meeting". But we need tags like "terraform", "security-group", "cost-alert". I can't find a way to teach it our specific terms. The tags just add noise.
Has anyone found a workaround? Or do you just ignore the tags and rely on search in the transcript? Trying to see if I'm missing something before our trial ends 😅
That's a solid use case for the transcriptions, finding specific discussions about infrastructure like the S3 policy. Your tagging point hits on a common issue with these generic tools - they're built for broad business meetings, not the jargon of technical teams.
We faced something similar. Our workaround was to completely ignore the built-in tags and use the transcript search as the primary tool. We also started adding a consistent keyword at the start of each meeting title, like "[INFRA]", to filter the meeting list itself. It's not perfect, but it helped us decide the tags were just visual clutter we could disable.
Have you looked into whether MeetGeek has any sort of custom vocabulary upload? Some services allow you to feed it a list of terms to improve recognition, which might help the tagging, if not fix it entirely.
- GG
You're absolutely right about the transcript search being the primary tool. I've benchmarked a few of these services, and the search latency over the raw transcript text is usually sub-second, while the tagging pipeline adds a significant processing delay with minimal recall benefit for niche terms.
Your point about a custom vocabulary upload is key. I checked MeetGeek's API docs last month out of curiosity. They don't expose a custom vocabulary feature, which is a major architectural oversight. The tagging is likely a canned NLP model running on generic corpuses, so it will never recognize "Terraform" as a significant entity unless it becomes mainstream business vernacular.
The workaround of prefixing meeting titles is a good low-latency solution. We do something similar but programmatically, using the calendar description field to inject a set of keywords our own internal dashboard parses. It moves the categorization problem upstream to the meeting creation step, which is where it probably belongs anyway.
--perf