AI voices on a phone call

Yes, synthetic speech can be detected in a recorded phone call — less reliably than in a clean file, and the gap is large enough that it should change how you read the result. Truthring reports [VERIFY]% on recorded calls against [VERIFY]% on studio-quality audio. This page explains where the difference goes.

What a phone network removes

A telephone call is not a recording of a voice. It is a reconstruction, built to be intelligible over a narrow channel at low cost. To achieve that, the network throws away almost everything that is not needed to understand words.

• Everything above roughly 3.4 kHz. Traditional telephony carries a band of about 300 Hz to 3.4 kHz. Human hearing goes to around 20 kHz. The upper region — where breath, sibilance and the fine texture of a real recording live — simply is not transmitted.
• The noise floor. Codecs are designed to suppress background, and modern ones aggressively remove what they judge to be non-speech. The traces of a real room and a real microphone are exactly what they are built to discard.
• Fine timing detail. Packet-based calls reassemble audio from fragments, conceal lost packets by interpolating, and re-time what arrives late. The result is intelligible and no longer a faithful record of what was said.

Put plainly: the phone network is a filter that strips out most of what distinguishes captured audio from manufactured audio, and it applies that filter equally to a real caller and a cloned one.

What survives, and what we measure

Not everything is lost. Detection on phone audio leans on the parts of speech that live inside the transmitted band and that a generator still has to get right.

How to read a verdict on a call recording

The asymmetry matters more here than anywhere else on this site.

• “Likely synthetic” on phone audio is meaningful. The evidence had to survive the codec to be found, so finding it is a real signal.
• “Likely human” on phone audio is weak. It can mean the speech was human. It can equally mean the evidence of synthesis was destroyed in transit. Do not treat it as clearance.
• “Unclear” is the honest outcome for many calls, particularly short ones. We would rather return it than manufacture confidence the audio does not support.

Detection rates by condition, including the rate at which real human speech is wrongly flagged: accuracy.

Getting the best recording you can

• Keep the original file. Every forward through a messaging app re-compresses it. Send us the file the recorder produced, not a copy of a copy.
• Submit the longest continuous stretch of speech, not the most incriminating sentence. Length is the single biggest lever you control on phone audio.
• Do not clean it up. Noise reduction and normalisation remove evidence along with the noise. Send it raw.
• Do not re-record it by playing it on a speaker and capturing it on another phone. That destroys the strongest remaining signal.
• Note the channel. A carrier call, a WhatsApp call and a Zoom recording compress differently, and telling us which one it was improves how we read the result.

What to do during the call, before any of this matters

No detector helps in the moment, ours included. The countermeasure that works is procedural and costs nothing.

Hang up and call back on a number you already have stored. Not a number the caller gives you. A voice clone cannot answer a phone you dialled, and essentially every voice fraud that succeeds does so because the target stayed inside the channel the attacker chose. Urgency is the tell: a real emergency survives a two-minute callback, and a manufactured one does not.

Questions