The smartest AI agent in the world still fails if it sounds bad. Callers decide in the first few seconds whether they're talking to something helpful or "a robot" — and on a phone call, AI voice call quality depends as much on how the audio travels as on what the AI says.
We listened to a lot of real ExoCall calls and found three separate problems. None of them was the AI model's fault; all three were in the plumbing between the AI and the phone line. Here's each one, in plain words.
Key Takeaway
ExoCall calls now sound clearer and more natural: a proper anti-aliasing filter removes the harsh, buzzy tones created when AI audio is squeezed onto an 8 kHz phone line; interrupting the agent clears its queued audio instantly; and short background noises like a TV no longer cut the agent off. The changes apply to every call automatically.
Problem 1: The Buzzy, Harsh Voice
The cause: the AI voice model speaks in high-quality 24 kHz audio. An Indian phone line carries 8 kHz audio, which physically can't hold any sound above 4 kHz. So the audio has to be converted — keeping one sample in every three.
The catch is that the high sounds don't just disappear when you do that. Unless they're removed first, they fold back into the range you can hear, as false tones. This effect is called aliasing. Our old converter simply averaged every three samples, which barely filters anything: a 5 kHz "s" sound came out only a little quieter — as a 3 kHz whistle. Multiply that by every s, sh and t in a sentence and you get the harsh, buzzy AI voice some callers heard.
The fix: a proper low-pass filter before converting — technically a 121-tap windowed-sinc filter. In practice:
Speech
Passed through untouched up to ~3 kHz
Above 4 kHz
Removed by 70+ dB before converting — no false tones
Added delay
2.5 milliseconds — far too short to hear
The filter also carries its state across every piece of audio the AI sends, so there are no tiny clicks where one chunk ends and the next begins. The result is the same voice you'd hear on a normal call — clear, without the metallic edge.
Problem 2: The Agent That Won't Stop Talking
The cause: real conversations have interruptions. A caller says "haan haan, samajh gaya" while the agent is still explaining. The AI model notices and stops — but the phone network had already received several seconds of its answer to play, and kept playing it. The caller heard the agent carry on talking over them, then suddenly answer something else.
The fix: the moment the caller interrupts, ExoCall now tells the phone line to throw away everything queued but not yet played. We also send audio in the smallest pieces the phone network allows (100 milliseconds each), so there's very little waiting in line at any moment. The agent now stops when you talk — like a person would.
Problem 3: The TV That Interrupts the Agent
The cause: fixing interruptions created a new problem. Indian callers often take calls with a TV on, family talking nearby, or traffic outside. The AI was treating those sounds as the caller speaking. On one test call, about half of 11 interruptions came from a TV or people in the room — and each one stopped the agent mid-sentence. With instant clearing, that sounded choppy.
The fix: we made the AI's "is someone speaking?" detector less jumpy:
- Lower start-of-speech sensitivity — short, faint or distant sounds are less likely to count as the caller talking.
- 300 ms of sustained speech — sound has to continue for about a third of a second before it counts as an interruption. A clatter or a TV line doesn't; a caller actually speaking does.
Background speech is also ignored when calls are scored afterwards: ExoCall's post-call analysis is told to disregard TV audio, other people in the room and stray fragments in unrelated languages, so a noisy room doesn't skew a lead's label.
Before and After
| Situation | Before | Now |
|---|---|---|
| Words with "s", "sh", "t" | Harsh, whistly, buzzy | Clean phone-quality voice |
| Caller interrupts | Agent keeps talking for a few seconds | Agent stops right away |
| TV or people in the background | Agent cut off mid-sentence | Agent carries on |
| Caller actually speaks | Interrupts | Still interrupts |
Tips for Clearer AI Calls
The engineering is on us. A few choices on your side help the conversation feel natural too:
- Keep the agent's replies short — 1 to 3 sentences. Long monologues invite interruptions.
- Pick a voice that suits your callers from the 10 voices in Settings, and test it on your own phone.
- Write numbers and prices the way they're spoken — "₹4,999 per month", phone numbers in groups — in your Business Knowledge.
- Make a test call from a noisy place, not just a quiet office. That's where your customers are.
Bottom Line
Callers don't care about sample rates or filters. They care whether the voice on the phone sounds clear and whether it listens when they talk. These fixes make ExoCall's AI agent sound less like a machine and behave more like a good receptionist — on real Indian phone lines, in real noisy rooms.
Live now on every call
No settings to change. Every inbound and outbound call on ExoCall uses the new audio pipeline.
Frequently Asked Questions
Why do AI voice agents sound robotic or buzzy on phone calls?
Usually because of how the audio is converted for the phone line. AI voice models produce high-quality audio (24 kHz), but a phone line carries only 8 kHz audio, which can't hold sounds above 4 kHz. If the high sounds aren't filtered out properly before converting, they fold back into the audible range as false tones — sharp "s" sounds turn into whistles, and the voice sounds harsh. ExoCall now uses a proper low-pass filter so this doesn't happen.
Can I interrupt an AI calling agent mid-sentence?
Yes. With ExoCall, when you start speaking over the agent, it stops and the audio already queued on the phone line is cleared immediately, so you don't keep hearing the rest of its old answer. The agent then listens to what you said and responds to that.
Does background noise interrupt the AI agent?
Much less than before. On one test call, about half of 11 interruptions were caused by a TV or people nearby, each cutting the agent off mid-sentence. ExoCall now needs about 300 ms of sustained speech before treating sound as an interruption, with lower sensitivity to short noises — so a TV in the background is ignored, but a caller who actually speaks still interrupts.
Do I need to change any settings to get the better call quality?
No. The improvements are on ExoCall's voice server and apply to every call on every workspace automatically.