Not reliably. Not in the way people mean when they ask.
I want to give the blunt answer first because this question usually arrives with a hopeful framing, and the hopeful framing is being actively encouraged by products that promise real-time deepfake detection on calls. The physics of telephone audio makes that a much weaker promise than it sounds, and the gap between the promise and the physics is where people get hurt.
Here is what is actually going on.
Detection evidence lives in the part phones throw away
A detector works by finding traces of the synthesis process, mostly left behind by the vocoder, the final stage that has to invent a waveform from an internal representation. Those traces are fine-grained spectral texture: the exact character of high-frequency noise, the consistency of harmonic structure, the phase relationships between components. I walk through the mechanics in how AI voice detection actually works.
Now consider what a phone network does to sound.
Traditional telephony was designed around intelligibility, not fidelity, and it was designed when bandwidth was expensive. The classic narrowband channel carries roughly 300 Hz to 3.4 kHz, sampled at 8 kHz. Everything above that ceiling is discarded before the audio reaches the other end.
A very large share of the evidence a detector wants lives above 3.4 kHz. The thin haze of turbulent airflow noise from teeth and lips, the high-frequency behaviour that vocoders historically reproduced carelessly, the fine detail in fricatives like s and f. Gone. Not degraded, not obscured. Removed at the source and never transmitted.
Wideband calling has improved this in places. Modern voice over IP and mobile codecs can carry more range, and a call carried end to end over a modern codec preserves more than the old copper-era channel did. But the improvement is uneven, it depends on both endpoints and every network in between, and you generally do not know which case you are in. A call that touches any legacy segment falls back to the lowest common denominator.
Compression removes what the ceiling did not
Bandwidth limiting is only the first cut. What survives then goes through a lossy codec.
Speech codecs are psychoacoustic. They model what human hearing does and does not notice, then aggressively discard the parts that fall below perceptual thresholds. That design goal happens to overlap almost exactly with the region where synthesis artifacts live. The codec is not trying to defeat detection. It just so happens that the low-amplitude fine structure a codec considers disposable is the same low-amplitude fine structure a detector considers evidence.
Modern speech codecs are also parametric in places, meaning they do not transmit the waveform at all but transmit parameters that let the receiver resynthesise something perceptually similar. Think about what that means for a detector. The audio arriving at the far end of the call was reconstructed by the codec's own synthesis stage. You are now looking for synthesis artifacts in audio that has definitionally been resynthesised. The signal you want and the noise you are fighting come from the same kind of process.
Add packet loss concealment on top. When packets drop, the receiver invents plausible audio to fill the gap. Again: invented audio, in a search for invented audio.
Then there are the platform layers
If the call is a conference rather than a plain phone call, another set of processing sits in the path. Noise suppression, echo cancellation, automatic gain control, and increasingly, machine learning based speech enhancement that reconstructs cleaner speech from a noisy input.
That last category deserves its own warning. Neural speech enhancement improves how a call sounds by regenerating parts of the signal. From a detector's point of view it is another synthesis stage applied to everyone, human and cloned alike, which pushes genuine speakers toward the artifact profile the detector is looking for.
This is a specific and underappreciated false-positive mechanism. Your genuinely human colleague, on a conferencing platform with aggressive enhancement, on a poor connection, has had their voice partially regenerated by a neural model before it reached your ear.
What our detector does and does not do
We built for files, deliberately.
Our voice detector takes an uploaded audio file, accepts MP3, WAV, M4A, OGG and FLAC up to 25 MB, and returns an AI score from 0 to 100 with the human score as its exact complement, plus a verdict of human, AI, or uncertain. There is no live call mode, and I am not planning one on the current evidence.
There is also a quality gate before scoring. If a clip is too short, too quiet, or does not contain enough actual speech, the detector declines to score it and returns uncertain rather than manufacturing a number. Phone recordings hit that gate more than any other category of file we see, which is not a coincidence. The gate is doing its job.
If you upload a recording of a phone call, you should expect weaker and more hedged results than you would get from a clean file. That is not the tool underperforming. It is the tool accurately representing how much evidence survived the journey.
An earlier version of ours got this wrong
Worth telling because it is the exact failure mode this article is warning about.
A previous model we ran, internally textsight:voice-v1, was scoring real human voice notes at 99 and 100 percent AI. Not occasionally. Routinely. The model turned out to be drawing a boundary between speech and music rather than between synthetic and genuine speech, and compressed close-miked phone audio landed on the wrong side of it.
Every one of those scores was confident. The confidence was real and the number was worthless. We pulled the model and replaced it with the anti-spoofing architecture we run now.
I bring it up whenever someone shows me a real-time call detection demo, because a demo cannot show you this class of error. A model that is confidently measuring the wrong thing looks exactly like a model that is confidently measuring the right thing, right up until you test it against messy real-world audio in volume. Ask any vendor how their tool behaves on ordinary human calls over bad connections, and treat a low false-positive claim without a published protocol as unsupported.
The asymmetry that makes false positives worse here
In file-based verification, a false positive is a cost you can absorb. You investigate, you find the clip is genuine, you move on.
On a live call, a false positive is an alert that fires while a real person is talking to you. Now consider the base rate. In a company where a handful of calls per year might be fraudulent out of hundreds of thousands, even a very low false-positive rate produces far more false alarms than true ones. That is not a criticism of any particular model, it is arithmetic, and it is the same reason rare-disease screening produces mostly false positives even with a good test.
An alert system that is wrong far more often than it is right gets ignored within weeks. Everyone in security knows this and it happens anyway.
So what actually protects you
The same answer as the finance controls piece, and it is boring on purpose.
Call back on a stored number. This is the single control that does not care how good the clone is. Attackers control what reaches you. They do not control where you dial. Hang up, look up the number from your own records rather than from the message, and call. If that is met with resistance, you have learned what you needed to learn.
Ask something not in any document. A voice model generates speech. It does not supply the speaker's memories. An attacker running a script has to improvise on specifics, and specifics are where improvisation shows.
Watch the request, not the voice. Urgency plus confidentiality plus a bypass of normal process is the signature. That pattern is far more diagnostic than anything in the waveform, and it is available to a person with no technology at all.
Preserve a recording. Where lawful and possible, keep the audio. This is where a detector becomes genuinely useful, as one signal in the incident file after the fact.
Notice that none of these require you to judge the audio. That is the point. Any defence that depends on a person or a model correctly evaluating how a voice sounds is a defence that voice cloning was built to defeat.
What about voice authentication, then?
A fair objection: banks and call centres have been using voice as an identity factor for years, over exactly these degraded channels. If detection is so hard on phone audio, how does that work?
The answer is that those systems are solving an easier problem, and it is worth seeing why.
Voice biometrics asks whether this speaker matches an enrolled voiceprint for this specific account. That is a comparison against a known reference, on a channel the operator controls, often with a known handset and a known network path. It is closed-set verification. Our problem, by contrast, is open-set: given arbitrary audio from an unknown source over an unknown path, decide whether it was machine-generated. Far less to work with.
Those systems do include anti-spoofing components, and the research lineage is shared. The ASVspoof challenge series that produced much of the modern countermeasure literature was built around protecting speaker verification specifically. But the countermeasure runs inside a controlled pipeline with a reference to compare against and an operator who can escalate to another factor when confidence drops.
The relevant lesson is what those systems do when the voice check is uncertain. They do not just wave the caller through with a warning. They fall back to a second factor: a code, a callback, an app confirmation, a question about a recent transaction. Voice is one input into a decision, never the decision itself.
Which is exactly the argument of this article, arrived at from the other direction by an industry that has been living with the problem longer than the media-verification world has.
When a phone recording is still worth analysing
I do not want to overstate the pessimism. There are cases where it is worth running the file.
A voicemail is often better material than a live call recording. It has been through fewer conversion steps and is frequently stored at a higher quality than what you heard.
A recording captured on the device itself, rather than pulled from a call platform, may have avoided a round of re-encoding.
A longer recording gives the model more to work with. Thirty seconds of continuous speech is a meaningfully different proposition from eight.
And a case where the audio was forwarded to you as a file rather than experienced as a call is often not really a phone problem at all. Plenty of what people describe as a suspicious call is actually a voice note, which is a file, and files are what detection is good at.
If you have something in that category, our voice detector is free to try without an account, and it will tell you honestly when it cannot judge.
The summary I would give a colleague
Detecting a cloned voice on a live phone call sits at the intersection of the two worst conditions for detection: the channel removes the high-frequency evidence, and the codec removes much of what is left. Products promising reliable real-time call detection are promising to work in the hardest available conditions, and the burden of proof for that claim should be much higher than a demo.
Meanwhile the control that would have stopped both of the well-documented cases costs nothing and does not depend on audio analysis at all. Hang up. Dial the number you already had. Ask a question the caller would need to actually be that person to answer.
Then, afterwards, run the recording. Just do not put it first.
Try it on your own writing