A few seconds of recording are now enough to imitate anyone's voice. Audio files carry less metadata than images and are easily degraded by calls and messengers, so checking relies heavily on detectors that read the sound itself.
Some generation services and open tools embed inaudible watermarks. Origma reads some public audio watermarks and checks the container and encoder record. A confirmed mark is strong evidence, but most circulating copies no longer carry one.
Synthetic speech leaves regularities in breath, noise and the spread of frequency bands that differ from human recordings. The box above shows the audio gate. By the owner's decision the audio threshold allows up to 3% of human recordings to be flagged, so it is more likely than images to wrongly flag a person.
Phone networks and messengers squeeze sound into a narrow band. Much of what the detector reads is lost on the way, so phone recordings come back inconclusive more often. The original file, if you have it, is far more accurate.
Do not wait for a score: first confirm with the person through another channel. The result only supports your judgement, and if money is requested, the rule is not to pay before confirming, whatever the result says.
This is general guidance, not legal advice. A scan result is a probabilistic assessment, not proof.