Detecting Bark

Most speech generators try to read a sentence cleanly. Bark, an open-weight audio generation model, does something else: it produces the sounds around the words too — a laugh, an intake of breath, a sigh, a false start, the noise a person makes while deciding what to say next. Those are precisely the cues listeners use to decide a recording is real, which makes this a useful case for explaining what our analysis does not depend on.

Why a laugh used to be good evidence

For most of the history of speech synthesis, the tell was flatness. Machines read sentences; people made a mess. Anyone assessing a suspicious recording could rely on a simple rule of thumb — if the speaker stumbles, restarts a word, laughs at their own sentence or breathes in the middle of a clause, a person was probably in the room.

That rule was never a measurement. It was an inference from what generators of the time happened to be bad at, and it held only for as long as they stayed bad at it. A model that produces those sounds directly does not defeat a detector. It defeats an intuition, and the intuition was load-bearing for a great many people who assess recordings informally: a manager listening to a voice note, a moderator reviewing a report, a family member deciding whether the call was real.

This is the practical significance of Bark, more than any particular quality of its voices. It removes the heuristic almost everyone was using without knowing they were using it.

What our analysis keys on instead

Truthring does not score how human a clip sounds. If it did, a laugh would count in the generator’s favour and this page would be an admission of defeat. The question we ask is narrower and duller: was this audio captured or was it manufactured? Those two histories leave different residue, and the residue is largely independent of which sounds the speaker makes.

The third of those deserves emphasis, because it inverts the expectation. A clip full of laughter and hesitation is not generally harder for us than a clip of clean read sentences. It is frequently a little easier. Ordinary declarative speech is what these models are trained on most heavily and reproduce most smoothly; a burst of laughter is an acoustic event with far more variability, generated from far less well-formed material. We report the two rates separately above rather than blending them, because the difference is the point.

Where this fails

• Very short clips built around a single vocalisation. A brief laugh with no speech around it gives the analysis almost nothing to work with. Confidence drops accordingly and the report says so.
• Clips assembled from several generations. Output stitched together in an editor, or interleaved with genuine recorded material, produces a mixed file. We analyse continuous stretches, and a short synthetic insert inside a real recording is among the hardest cases we handle.
• Re-recording through a speaker. Playing a generated clip aloud and capturing it on a phone gives the file a genuine capture history and removes the first pass from the equation.
• Heavy compression. Messaging apps discard the fine structure the analysis reads. This is the general limitation on every generator, and it applies here in full: see phone and compressed audio.

What to tell people who assess recordings by ear

If your team has an informal rule about what a real voice sounds like, this is the generator to demonstrate it against. Not because Bark is the system most likely to appear in a fraud attempt — it is not — but because a single clip of convincing generated laughter retires the rule faster than any amount of explanation.

Replace the rule with a procedure. The question is never “did that sound real”; it is “did this reach me through a channel I chose, and can I confirm it through one I control”. Ring back on the number you already had. Ask something only the person would know, and be aware that public material makes a poor test. Keep original files rather than forwarded ones, because every re-send strips evidence.

Questions