# Call quality The gap between a caller finishing and the agent answering, the two settings that shape the line, and the trade-off between them. Source: https://docs.omazy.ai/how-to/voice/call-quality/ import Figure from '../../../../components/Figure.astro' import { Aside } from '@astrojs/starlight/components' A voice agent is judged on timing before it is judged on answers. A correct reply that arrives after three seconds of dead air loses to a vaguer one that arrives immediately. ## The gap, and what fills it Between a caller finishing their sentence and the agent starting to speak, several things have to happen: notice they stopped, turn the audio into words, work out an answer, and turn that answer back into speech. None of it is instant.

This is why [acknowledgements](/how-to/voice/phrases/) are the highest-value phrases you can write. They do not make anything faster. They change dead air into something that sounds like a person thinking, which is the part the caller actually reacts to. ## The two settings Both live on the **Voice** tab. Both are trade-offs, and the defaults are reasonable, so change them only against a specific complaint. ### Pacing lead The agent's speech is sent slightly ahead of when the caller needs to hear it. That small buffer absorbs network hiccups. With no lead, at most a fraction of a second of audio is ever ahead of the caller's ear, so any stall between the platform and the phone network is heard as a gap in the middle of a word. Adding lead smooths those out. The cost is that interrupting the agent takes marginally longer to take effect, because there is more audio already in flight to discard. **Raise it if** callers report choppy or stuttering speech. **Leave it alone if** the line already sounds smooth. ### Echo suppression If the caller is on a speakerphone, the agent's own voice comes back through their microphone. Without protection, the agent hears itself, treats it as the caller speaking, and answers its own sentence. It is unmistakable once you have heard it: the agent starts talking to itself and the caller cannot get a word in. Echo suppression stops the agent listening while its own audio is still reaching the caller, plus a short tail for the room. ## What good sounds like - **The greeting starts immediately.** Any pause before the first word is the worst pause on the call, because the caller does not yet know anyone is there. - **Replies begin within a beat or two.** Some questions take longer, which is what acknowledgements are for. - **Interrupting works**, unless you have deliberately traded it away. - **No clipped word endings.** Speech cut off at the end usually means the agent decided the caller had stopped talking too eagerly. ## Things outside these settings Some call quality problems are not settings problems, and recognising them saves a lot of tuning. **The caller's own connection.** A caller on a poor mobile signal will hear a poor call regardless. Their audio reaching the platform intact matters as much as anything on this page. **Distance.** A call routed through a distant part of the network before it reaches the agent carries that round trip on every single turn. If test calls from one location are consistently slower than another, that is usually why, and no amount of pacing adjustment fixes it. **Speakerphones in hard rooms.** A tiled room with a speakerphone is the worst case for both echo and recognition. It is worth testing one deliberately, since a meaningful share of real calls arrive that way. ## Testing changes Change one setting at a time and place a real call after each. These settings interact, and two changes at once leaves you unable to attribute the result. Test on the kind of connection your callers actually use. A change that improves a call from a quiet office over good WiFi can make things worse for someone on a phone in a car, which is where many of your calls come from.