Recordings are common evidence in disputes, but a few seconds of someone's voice is now enough to produce a convincing AI copy. Whether a recording was cut or spliced and whether the voice itself is synthetic are separate questions that need separate evidence.
Recording apps and devices leave the file format, sample rate, codec settings and the name of the software that wrote the file. Editing software changes these, and at a splice the background noise or loudness can jump. Origma reads this information at the metadata stage and records where the file departs from an original recording.
AI-generated voices tend to differ from real recordings in breathing, transitions between sounds and frequency distribution. Origma's audio detector turns these traits into a score, and when a watermark that some generation services embed is readable, that is recorded too. Thresholds are set conservatively to avoid calling real voices AI.
Phone calls strip a lot of audio quality in transmission, and voice messages are recompressed by the messenger. Subtle traces disappear along the way, which makes detection harder. Read call recordings and forwarded audio with that in mind; when we cannot tell, the result is inconclusive.
Origma judges edit traces and AI-generation likelihood separately, but it does not identify who is speaking. Results are probabilistic assessments, not certified forensic reports or legal proof. In an important dispute, keep the original recording file and get an expert examination as well.
This is general guidance, not legal advice. A scan result is a probabilistic assessment, not proof.