Your voice agent is talking to someone watching TV. How bad could it be?

This is an extension of VAmoS Pro Bench, our voice-agent benchmark (here is the launch post). A television in the caller's room fails most calls. Failing the task is only part of the damage. The agent also mishears the caller, talks over them, cuts itself off, and runs up four times the bill. We ran one agent through 100 tasks, three calls per task, both on its own and with ai-coustics' Quail Voice Focus model in front of it. With the filter, TV calls looked like quiet ones on every measure.

Share of calls completed with each background (a quiet room, street noise, café chatter, or a TV), 74 to 75 calls each. The TV difference is significant (p = 0.0001, Fisher's exact test).

27 seconds of one call without the filter, at phone quality (8 kHz). The caller is trying to switch to paperless billing.

Listen for the TV. You will probably need headphones, because it plays about 25 dB below the caller's voice. Here is what the agent heard in those 27 seconds and how it answered.

It heard “So many problems. Then that guy” and replied about the payment it had just recorded.

It heard “you could still film him” and replied “I'm here to help with your account.”

It heard “But he's, you know, you can't upload” and replied “If you have questions about your bill or payment plan…”

Meanwhile the caller started “I want to switch…” six times and was talked over five times.

No person on the line would have noticed that TV, let alone answered it. A person hears the caller close to the phone and a voice across the room and ignores the second one without trying. This agent, like most voice agents, turns the audio into text and hands the text to a language model. Once the audio is text, the presenter's words look exactly like the caller's.

	Quiet room	TV	TV + Quail Voice Focus
Calls completed	20%	1%	21%
Transcription error	10%	154%	13%
Caller turns the agent talked over	33%	63%	31%
Agent turns cut off midway	29%	60%	28%
Times the agent spoke, per call	12	42	10
Call length	3.0 min	4.4 min	2.7 min
Cost per call	$0.14	$0.52	$0.12

The Pipecat agent on VAmoS Pro Bench, 75 calls per column, from the same 100 tasks and simulated callers in both runs. Transcription error and times the agent spoke are medians per call, and the rest are over all calls. Every difference between the two TV columns is significant (p < 0.002). Cost is the agent's own usage at list prices for gpt-4.1-mini, Deepgram, and ElevenLabs. Quail Voice Focus adds $0.00135 per minute on ai-coustics' Startup plan ($135 a month for 100,000 minutes), which is under half a cent per call.

Noise does more harm than a failed task

Most TV has people talking on it, and ours was a news show. Every part of a voice agent is built to react to speech, so one extra voice sets off a chain of failures.

First, the agent hears the TV as the caller. The speech recognizer writes the broadcast into the caller's turns, and on TV calls it made more errors than the caller said words. The word error rate was 154%, against 10% in a quiet room.

Then it interrupts in both directions. The agent can't tell when the caller has stopped, so it talks over them on 63% of their turns. When the TV talks while the agent is speaking, the agent takes it for the caller and stops. It cut itself off on 60% of its turns.

It also answers the TV. On a typical TV call the agent spoke 42 times, against 12 in a quiet room, mostly replying to things the caller never said. Calls ran 50% longer.

You pay for all of it. Every reply is generated, spoken, and billed, even the ones cut off halfway. A TV call cost $0.52 against $0.14 for a quiet one, and most of the difference was text-to-speech.

Speech enhancement as a drop-in

Removing traffic and hum is routine. Removing a second voice while keeping the first is harder because both are speech. Quail Voice Focus is a speech-enhancement model built for that job. It keeps the main speaker and removes everything else, other voices included. Pipecat ships it as a built-in audio filter, so adding it is one setting. It added 30 ms of delay.

Does it work for your agent?

Whether Quail Voice Focus helps your agent depends on your speech recognizer, the noise your callers deal with, and what the agent does with what it hears. You find out the way you would pick a model or an architecture, with a benchmark.

We took the Pipecat agent from VAmoS Pro Bench and ran it twice. Each run had the same 100 tasks, three calls per task, with the same simulated callers. Every call had one of four backgrounds (a quiet room, street noise, café chatter, or a TV), and each background covered a quarter of the calls in both runs. One run had no filter. The other ran Quail Voice Focus (quail-vf-2.2-l-16khz) on the caller's audio. Nothing else changed, and we ran both at the same time.

What changed and what did not

TV calls recovered on every measure. Completion, transcription, interruptions in both directions, call length, and cost all came back to about where a quiet room was. Across all 300 calls, 27% completed against 17%, and the average call cost $0.12 against $0.23. The filter did no harm on easier calls either. In a quiet room and on the street, completion and transcription stayed the same.

What this does not show

We tested one agent with one speech recognizer (Deepgram nova-3) and Quail Voice Focus at its default settings. The TV was one recording mixed into the caller's audio, not a real room. Your costs depend on your providers. Here most of the difference came from text-to-speech billed per character, and the per-minute price of Quail Voice Focus assumes you use the minutes on your plan. Quail Voice Focus also does not fix turn detection. In a quiet room this agent still talks over a third of caller turns, which is the problem Jev addresses.

Run it on your own calls

These results come from 600 simulated calls. Veris can build the same kind of simulation around your own callers, policies, and noise, so you see what a new component does to your calls and your bill before customers hear it.