Here's a pattern I keep running into. Someone builds a voice agent, talks to it from their laptop for an afternoon, and it's great. It catches every date and reads back every order number in a warm, friendly voice. Then it goes live on a phone number and the first week of calls is full of misheard digits and an agent that suddenly sounds tired.
What changed was the audio reaching the model.
The browser hears a cleaner world
When you test from the browser, your voice reaches the speech-to-text model through a decent microphone at 16 kHz. A traditional phone call is narrowband: 8 kHz, and then squeezed by the phone network's codec on top of that. Which one your agent gets also depends on the telephony provider. In Dograh, for example, Twilio, Telnyx, Plivo and Asterisk calls arrive at 8 kHz, while Vonage calls arrive at 16 kHz.
So the "same agent" can be listening to two very different signals depending on how the caller dialled in.
The codec is often the bigger problem than the sample rate. A June 2026 study from the Indian Institute of Science found that, on Hindi speech, plain narrowband audio kept accuracy close to the original, while older compressed codecs like GSM hurt it consistently.
How bad it gets depends on the model
This is the part that surprises people. On WildASR's March 2026 telephony tests, six of seven modern models barely moved on English read speech. Whisper Large V3 went from 4.2% word error rate to 4.8% on a standard phone codec. One model, Qwen2-Audio, went from 5.8% to 48.3%.
Real phone paths can be rougher than one clean codec. Caller Digital ran a vendor benchmark across five Indian languages in September 2026: 7.4% word error rate on 16 kHz audio, 11.8% on a normal 8 kHz line, and 19.7% when a landline call ended on a mobile. Long digit strings came out fully right 92.8% of the time on 16 kHz and 81.4% on the phone line.
In our own work we've seen phone-call transcription errors run up to 3x higher than web calls. That number swings a lot with the speech model and the route the call takes, which is exactly why you have to measure your own.
The words that break first are the ones your flow leans on hardest, like account numbers and dates. And the failure is quiet. A misheard digit sends the agent to the wrong customer record, and it reads that record back with total confidence.
The voice suffers too
Narrowband audio strips the high frequencies that make a synthetic voice sound natural. In the same Caller Digital benchmark, a Hindi listening panel gave a premium neural voice 4.4 out of 5 on wideband audio and 3.6 on a normal phone line. The gap between premium and mid-tier voices shrank too, so the voice you paid extra for may not earn it over a phone.
My rule now is simple. Pick the voice by listening on an actual handset, never through headphones.
Mobile calls help, until they don't
Plenty of mobile calls today run at 16 kHz or better, thanks to VoLTE and 5G calling when both callers are on compatible carriers. The catch is that the codec behind it is designed to fall back. When the signal drops, or the call crosses into a landline, you're back on narrowband in the middle of a sentence.
What to actually test
Test the worst path, because that's the one your callers will find. Call the agent from a landline. Call it from a mobile standing somewhere with one bar. Read out a long account number and a date with an awkward month, and check the transcript against what you said. Then listen to the agent's voice on the handset, not on your laptop.
If you're on Dograh, an open-source, self-hostable voice agent platform, the run's trace shows exactly what speech-to-text heard on each turn, so a misheard digit is easy to tell apart from a confused model. We also wrote up the full order we test and scale agents in, from simulated callers to a slow ramp under load, with the data behind each step. Read the full guide on Dograh if you want the rest.
A real phone test is the cheapest check in the whole process, and in my experience it's the one teams skip most.
Top comments (0)