Why does it differ so much? Call recording has beautiful voices however actual phone calls are not that pretty. I did 3,278 real call so far and tbh none of them sounded anything near that is on the recordings. It cannot be network or telephony provider(Twillio/11lab/cartesia) or fault in general as an excuse. It more like Instagram beauty filter on the voice recording, when actual person and image on the screen are absolute different if you know what I mean. My boss heard recording and he said is awesome but now I am told - not doing the job because it doesn't sound anything that he sign off when he decided to move from Bland to Retell based on call samples (recordings) 🙂 so who's fault it is selling dream calls ? 🙂

hey @Q4 Recordings are higher quality than live phone calls by design. So this happens Because: -Call recordings capture the voice before it goes through the phone network -Real phone calls use PSTN compression (8kHz, μ-law), which removes a lot of detail this makes live calls sound flatter and less natural than the recording. so: -Recording = near-raw TTS -Live call = compressed phone audio that’s why the difference feels big and why your boss hears something nicer in the recording. What you can do is : -Pick voices that hold up well at 8kHz -Slightly slow the voice and reduce expressiveness

However, the industry hype focuses on how these models sound in demos and their ability to realistically mimic human voices. It would seem more logical to evaluate them against genuine recordings, rather than high-fidelity audio that is never actually encountered in phone conversations.

you’re absolutely right and that’s a fair criticism. most demos showcase studio-quality TTS, because that’s where the models sound the most impressive. but PSTN phone calls are a very different medium, and the gap isn’t always communicated clearly enough. In reality Phone calls never use that high-fidelity audio. Everything is downsampled and compressed and Some voices that sound amazing in demos degrade badly on real calls. we agree this expectation gap exists, and feedback like this is valid.