I've been evaluating Otter.ai for automating meeting note transcription and structuring within our analytics team's data pipeline. While the speaker identification and overall accuracy are impressive for our technical discussions, I've encountered a consistent formatting issue: Otter's handling of enumerated lists and bullet points is highly inconsistent.
The problem manifests in several ways:
* Lists are often transcribed as run-on sentences: "We need to one fix the data latency issue two update the dashboard filters and three review the new schema."
* Explicitly spoken markers like "bullet point" or "dash" are frequently included verbatim in the transcript, breaking the intended structure.
* Numbered items are sometimes correctly formatted but then revert to plain paragraphs within the same conversation.
This creates a significant post-processing burden. We feed these transcripts into a downstream system that parses structured action items, and the lack of reliable delimiters breaks our automation.
Has anyone developed a reliable workflow or set of prompting strategies to coax better list formatting from Otter? I'm particularly interested in:
* Whether a specific speaking cadence or phrasing (e.g., saying "new line" between items) yields better results.
* If there are hidden settings in the enterprise/admin panel that govern formatting logic.
* Any post-processing scripts you've used to remediate this. A regex-based approach is my fallback, but it's brittle.
Example of a problematic transcription output:
```
We have three priorities: first we need to validate the overnight batch job results bullet point second we should escalate the API timeout issue and finally third don't forget the stakeholder sync on Friday.
```
My ideal output would be structured data, even if simple:
```
1. Validate the overnight batch job results.
2. Escalate the API timeout issue.
3. Don't forget the stakeholder sync on Friday.
```
Any insights from your own data quality efforts with this tool would be appreciated.
Data is the only truth.
Oh man, I feel this pain. I've hit the exact same wall with Otter's list formatting, especially when speakers get into rapid-fire technical items. The verbatim "bullet point" inclusion drives me nuts!
One weird trick that gave me mixed but slightly better results was experimenting with *how* I speak the list markers during a test recording. Instead of saying "Number one," I tried a very deliberate pause before and after the number, like "...first item [pause] one [pause] fix the latency." It's awkward, but Otter sometimes interpreted that as a line break. It's not a real solution, though, just a hack.
I'm starting to think the post-processing route is unavoidable. Have you looked at throwing the raw transcript through a lightweight parser, maybe using a regex to catch patterns like "one X two Y three Z" and reformat? It's extra work, but maybe less than manually editing every transcript. What downstream system are you feeding into?