Why does the AI understand office callers fine but struggle with people calling from a jobsite?
The audio itself is worse before any AI touches it. Phone calls are compressed to a narrow frequency band, mobile calls lose packets, wind and handling noise hit the microphone directly, and truck speakerphones add echo. Recognition degrades with all of it. The diagnostic question is whether errors are spread evenly across all calls, which points at configuration, or concentrated on mobile and speakerphone calls, which points at the channel.
What the phone network does to speech
Telephone audio is band-limited. A large part of the frequency range that distinguishes consonants like s, f and th simply is not transmitted. Humans compensate using context so effectively that most people never notice. A recognizer compensates less well, which is why misrecognitions on phone calls cluster on short words and word endings rather than on long distinctive ones.
Mobile adds packet loss and variable codecs. A call that sounds acceptable to a person can be missing short slices of audio, and each missing slice becomes a guess.
The field-specific offenders
- Wind across the microphone. Produces low-frequency energy that noise suppression fights, often by removing speech along with it.
- Truck speakerphone. Echo plus road noise plus distance from the microphone. The single worst common case.
- Handling noise. A phone against a shoulder while someone works produces bursts that the endpoint detector reads as speech.
- Competing voices. A second person talking on site defeats speaker separation more often than vendors admit.
Aggressive noise suppression is not free
The reflex is to turn noise reduction up. Beyond a moderate setting, suppression starts removing the quieter parts of speech along with the noise, and the recognizer receives a cleaner-sounding signal with pieces of words missing. That produces confident wrong transcripts, which are worse operationally than obviously garbled ones, because downstream automation acts on them.
Tune it against your own recordings and measure the outcome, not the perceived cleanliness of the audio. The right measure is the downstream one: did the address end up correct, did the job get booked correctly, did anyone have to call back.
How to localize the problem
Segment your call data by inbound number, by whether the caller was on a mobile or landline where that is available, and by hour of day. Then compare repeat rate, transfer rate and correction rate across the segments.
Even spread across all segments means the problem is upstream of the channel: a prompt, a vocabulary, a model version. Concentration in mobile calls during business hours means it is the channel, and the fixes are different, mostly shorter prompts, more confirmation, earlier keypad fallback and faster escalation. Getting that segmentation requires joining the phone system's metadata to the transcripts, which is a small piece of integration that pays for itself the first time it prevents a pointless prompt rewrite. Call intelligence is where the joined view lives.
Topics: audio quality · speech recognition · mobile · diagnostics · field
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.