Hi everyone — I’m new here, and honestly a bit out of my depth with AI tools. I’ve been tasked with evaluating Read AI for our team’s migration planning meetings. We often do video calls with screen shares to walk through legacy database schemas and AWS console setups.
The problem is, whenever someone shares their screen during the call, the transcript gets really garbled. It’s like it tries to transcribe the text or UI elements on the shared screen, but mixes it up with our spoken conversation. We end up with a jumble of technical terms, random numbers from diagrams, and our actual dialogue all mashed together. It makes the meeting notes almost useless for our post-meeting action items 😅.
Has anyone else run into this? I’m hoping there’s a setting or a best practice we’re missing. We’re using Zoom for the calls, with Read AI joining as a participant.
What’s the step-by-step to fix this? And realistically, how long did it take you to get clean transcripts with screen sharing? We’re trying to lock down our tooling for the cloud migration, and I need to report back on whether this is a reliable solution.
One step at a time
Oh man, this is a classic "garbage in, garbage out" AI situation. The tool is trying to listen *and* read everything at once, and it's not smart enough to separate your voice from the text in a screenshot.
You're not missing a setting in Read AI, you're missing a setting in your meeting habits. Stop letting the AI "see" the screen share. The transcript is coming from the audio stream, but the garbled text is coming from the visual OCR of the share. You need to force it to only hear the audio.
In Zoom, when Read AI joins as a participant, it gets both the video feed (your shared screen) and the audio. Try this: In your Read AI meeting settings, configure it to only process the audio from the main Zoom meeting, not the combined audio/video feed. Sometimes this is called "audio-only mode" for transcription.
If that's not an option, a low-tech workaround is to have the person sharing screen briefly stop sharing when there's a critical verbal discussion that needs a clean transcript. I know it's clunky, but it's the only thing that worked reliably for us.
Realistically, it took us two weeks of messed up notes before we figured this out. Now our transcripts are about 90% cleaner. Still get the occasional stray "us-east-1a" popping into the conversation, but it's manageable.
It's always DNS.
Forget the fix. You're using the wrong tool for the job. Read AI is designed for basic sales calls, not parsing complex technical screenshares with database schemas.
Even if you force audio-only mode, you're losing all the context from the screen. The transcript becomes useless for your migration notes because it can't reference the diagrams you're discussing.
This is a fundamental limitation. I'd scrap it and look at something built for engineering workflows, or just record the meeting and take manual notes. These AI transcript tools fail when you actually need them.
Don't panic, have a rollback plan.
Yeah, we get the same thing in HubSpot meetings sometimes. I think user338's idea about forcing audio-only is the right track.
But honestly, if your diagrams are important, maybe use a different note-taking method just for those parts of the meeting? Like have one person live-edit a doc with screenshots. Then use the transcript for the general talk parts.
How many people are in these meetings? I wonder if fewer participants makes the mixing less likely.
Trying to figure it out.
I think you're right about the tool's limitations for pure engineering workflows, but I'd push back on scrapping it entirely. The key is managing inputs.
>you're losing all the context from the screen
That's true if you rely solely on the AI output. But we've used a workaround: run the audio-only transcript in parallel with a simple screen recording. The clean transcript timestamps the conversation, and you can jump to the matching moment in the video when someone says "see this schema here." It's not automated, but it's way better than manual notes for the whole call.
These tools need a very specific scope to work. They fail when we expect them to do everything.