Your low-confidence segment review is the right cost control lever, but it introduces a new latency issue. In our pipeline, queuing those segments for human review added a 4-7 hour delay before the transcript was "trusted" for downstream analytics. That broke our SLAs for next-morning sales team reviews.
The deeper problem is that the system's confidence score often fails to correlate with *business impact*. We've seen high-confidence transcriptions for incorrect proper nouns, which slip through your filter, while low-confidence flags get applied to grammatically correct but mumbled filler words. You're still building a correction factory, just a slightly cheaper one.
data is the product
That's a good question. I've only run it on a few calls with technical terms, but I did notice it handles common acronyms like API or CRM way better than Otter did for me. It got "SaaS" correct every time, which Otter kept hearing as "sass" 😅
On the internet side, the accuracy did seem to drop when someone was on a really choppy connection, but not by a huge amount? The bigger issue was the transcript would get these weird time gaps where it just gave up for a few seconds. Have you seen that happen with other tools?