Audio case study · October 27, 2026

Real vs AI Voice Examples

Listen to two human recordings and two AI-generated English clips, then compare their known origins with the recorded detection results.

Both AI clips returned Likely AI. Both human recordings returned Unclear after a score interpretation fix. These four samples are a small case study, not an accuracy estimate.

Listen to the human recordings

The site owner supplied these two recordings and identified them as their own real speech recorded in Mac Voice Memos. These are demonstration recordings, not recordings collected from customer uploads. The original files use mono AAC audio at 48 kHz. Location-based filenames have been replaced with neutral sample names.

Listen to the AI-generated clips

Both clips use Kitten TTS through OfflineTTS with the Hazel preset, 1.0x speed, and WebGPU processing. They were exported as mono 24 kHz PCM WAV. These are preset synthetic voices, not voice clones. The scripts below are the text entered for synthesis, rather than independently verified transcripts of the human files.

What were the actual detection results?

Recorded October 27, 2026. Known origin describes how each file was made. AI score is the clip-level score displayed using our classification rules; it is not a measured accuracy rate or proof of identity.

Likely Human: score at or below 30%. Unclear: strictly between 30% and 70%. Likely AI: at or above 70%. A low AI score is not a guarantee that a message is genuine or safe.

Why did the real recordings return Unclear?

The human recordings contained mixed segment-level signals. For recording 1, two speech segments were labeled non-synthetic and one was labeled synthetic; a final no-content segment was excluded. For recording 2, three segments were labeled non-synthetic and one was labeled synthetic. Converting each valid segment to the same AI-score direction and averaging equally produced 42.12% and 36.45%.

These are inconclusive results, not successful human identifications. We do not know which recording characteristics caused the mixed signals. An unclear result should prompt source verification rather than an accusation. We report both uncertain cases instead of showing only the two AI clips that matched their known source.

A scoring issue this case helped uncover

During review, we found that our score interpretation code incorrectly treated the non-synthetic label as synthetic because it matched part of the text. It also counted no-content segments as evidence. We fixed both behaviors and added regression tests. The human scores above were recalculated from the original segment responses using the corrected code; they are not fresh detection requests.

The correction fixes how evidence is interpreted. It does not prove the underlying detection is always right. The two AI clip scores were unchanged. The confidence label is based on distance from the score midpoint, rather than a separate validation of correctness.

Test conditions and reproducibility

The everyday AI clip was submitted as its original 24 kHz WAV through the public detection route with a 15-second analysis cap; the full file is 12.925 seconds long. The work-update AI clip and both human files were tested internally in their published formats, as full files, without the browser upload conversion. The internal path used the same score interpretation code but did not consume public user allowances.

These paths differ from a browser upload that converts audio before detection. We did not run a controlled same-format, same-speaker comparison, and do not claim that every future upload will produce these scores. The exact model build was not exposed in the saved results. No additional calls were made to search for a more favorable result. The sample manifest records file hashes, scores, and how each result was obtained.

How were these samples prepared?

The human files retain their original audio encoding; recording-level container metadata was removed for privacy. The generated WAV files were exported directly from the synthesis tool. We applied no additional noise removal, music, pitch changes, or speed edits to these published samples. The different voices, recording conditions, and file formats mean this is not a controlled experiment that isolates only the effect of synthesis.

The generation interface exposed the model family and settings, but not an exact model build identifier. File sizes and SHA-256 checksums are available in the sample manifest. Keeping source information separate from a detector result makes it clear what is known and what has not been measured.

What should you listen for?

Listen to pauses, sentence endings, breath sounds, emphasis, and transitions between words. These features can help describe a clip, but no single audible characteristic reliably proves that speech is synthetic. Human speech may sound unusually steady, and generated speech may contain breaths, hesitation, or expressive changes.

Try both passages instead of choosing only the phrase that sounds most artificial. An unfamiliar accent, a stumble, or polished delivery is not proof of generation. Recording hardware and compression can also change what you hear. Because the human speaker and the generated preset are different voices, differences between them cannot all be attributed to synthesis.

What can you conclude from four clips?

These files provide a small, accessible listening reference with stated origins. They do not establish detector performance on other speakers, generators, voice clones, languages, phone calls, or noisy recordings. Two AI samples matched the Likely AI category; two human samples remained Unclear. This is too small and too narrow a set to establish false-positive rates, false-negative rates, or overall accuracy.

For a suspicious message, confirm the sender through a contact method you already trust. For a podcast or video, ask for the original recording and disclosure of synthetic narration. Read our accuracy and result interpretation guide before treating a score as evidence. You can also return to the AI Voice Detector to check audio you have permission to submit.

Voice example questions