We use Deepgram's streaming API as the ears of an in-house AI phone agent that answers calls for a car transport business. It takes the raw 8 kHz phone audio straight off the line, no converting, and turns it round in under a second, which is what makes a live conversation feel natural. The API is simple to wire up, it's been reliable in production, and it's a small part of our cost per minute. When we compared nova-3 with the older phone-call model on 24 real recorded calls, nova-3 got noticeably more Australian town names right (Kalgoorlie, Kempsey, Archerfield) at the same speed and a lower price. Review collected by and hosted on G2.com.

Short words on a phone line are still hit and miss. "Quote" comes out as "card" or "quarter" every so often, on both models. There's no setting to ignore a second voice or background noise in the caller's room. Moving to nova-3 also meant changing how we pass hint words, because it rejects the old keywords option. That was easy enough, but it would have been nice to see it flagged more clearly. And speaker labels on streamed phone calls haven't been useful to us. Review collected by and hosted on G2.com.