Yeah, that's the part I'm stuck on too. So you're saying even after you get the words in order, you still have to manually build a system to find "objection handling" in them? That's a huge lift 😅
What if you just flagged words from a simple list instead of training a classifier? Like, scan for "competitor" or "too expensive" and tag the whole phrase? Or is that not accurate enough?
Containers are magic, but I want to know how the magic works.
That's exactly the breakthrough moment! The JSON structure with phrases is the key to moving beyond just a text dump. I've been feeding those `speaker_label` and `phrases` arrays into a simple script that aligns topics with our CRM's deal stages.
One caveat I hit early: speaker labels reset for each transcript. If you're tracking a specific AE's performance across calls, you'll need a mapping step to tie "Speaker 1" back to a real person in your CRM, otherwise the BI data gets messy.
But once you've got that mapped, seeing the correlation between how often "security" comes up in early calls and the deal velocity is super revealing. It's like unlocking a whole new data layer.
Let the machines do the grunt work
You're right about the JSON being usable, but the plain text export is useless. The structure is what matters.
I benchmark API response times. Otter's `phrases` array extraction adds 10-15ms per API call versus grabbing the raw transcript text. It's negligible until you scale to hundreds of calls.
The real bottleneck is mapping topics. You'll need to run the phrase text through a separate model for "pricing" or "security" detection, which is where your latency will spike. Their API just gives you words in order.
Benchmarks don't lie.