I've been using Read AI for my team's standup meeting summaries. There's a bug where it often mislabels who is speaking, especially when two people talk quickly.
Support hasn't responded in 5 days. Has anyone found a workaround? I'm wondering if adjusting the audio input settings or using a different recording format helps. I'm currently using the standard web app upload.
learning every day
The audio format is a good hypothesis. We've seen similar issues with other diarization services where low-bitrate compressed formats introduce artifacts that confuse voice activity detection.
A more reliable interim workaround, albeit more manual, is to pre-process the audio file yourself. Use a tool like Audacity or ffmpeg to split the audio into separate channels or files if each speaker is on a distinct input source, like a dedicated microphone. Uploading isolated streams can force the system to assign speaker labels correctly based on the source file.
Have you tried using a lossless format like WAV or FLAC for your upload instead of a compressed MP3/M4A? The larger file size might be worth the accuracy gain.
Show me the bill.
That's a great point about the upload format. I'd been using the default .m4a from my recording app, so switching to WAV is an easy test.
A quick follow-up question though - when you mention splitting into separate streams, are you assuming each speaker is on a separate mic? In our standups, everyone's on one conference mic. Is there any point in splitting a single source audio file, or does that just make the same problem happen in multiple files?
Thanks!
Yeah, I've run into this exact problem during our project syncs. Support was slow for me too, actually.
When two people start talking over each other even a little, the labels get completely swapped. Have you noticed if it's worse with remote callers vs. people in the same room? I'm trying to isolate the cause.
I haven't tried different audio settings yet, but now I'm wondering if using a better external mic for recording would make a difference, or if the bug is purely in the software processing.