← AutoDirector

Voiceprints vs audio levels:
who's talking on a shared microphone

Measured on a three-person dialogue testbench · identical material, identical scoring · all raw data published

Most "AI camera switchers" decide who is talking by comparing audio levels per microphone. That works when every person has their own mic. But plenty of real podcasts run on a single shared microphone — one USB mic in the middle of the table, one channel, everyone on it. Levels-based switching has no information to work with there, and this is measurable.

The testbench

Three speakers, one shared microphone, scripted dialogue with realistic crosstalk (people interrupting, laughing over each other, back-channeling "mm-hm"). The audio was played into OBS through the same path a real stream would use, and we scored every camera cut against ground truth: was the shot on the person actually speaking?

Results

ConditionLevels-based switcher ($69 market leader)AutoDirector (voiceprints)
Clean audio, one mic per person100% correct100% correct
Crosstalk −12 dB80%100%
Crosstalk −3 dB27% of turns86–100% of turns
One shared mic (three people) 0% — cut wrong every time holds the wide shot; voiceprints decide close-ups

The levels-based product doesn't just degrade on a shared mic — it confidently cuts to whoever isn't talking. On the shared-mic condition it produced 0 correct shots while AutoDirector held the wide shot instead of guessing, and switched correctly once voiceprints were enrolled (22 consecutive cuts, 100% attribution on a two-person validation run).

Why levels can't work on a shared mic

When everyone shares one channel, "who is louder" is undefined — there is only one level. A switcher can detect that someone is speaking, but not who. Any product that claims level-based speaker switching on a shared mic is guessing, and the guess lands on the wrong person as often as not.

Voiceprints are different: each speaker enrolls a short voice sample once. After that the switcher matches live audio against those samples and knows which voice is speaking — no extra hardware, still one microphone.

What we refuse to do

When the app can't tell who is talking, it holds the wide shot instead of cutting to the wrong person. A wrong close-up is worse than a correct wide. That restraint is a feature, and it's the difference between "usable on day one" and "embarrassing on day one".

Try it on your own show

AutoDirector is a Windows app that drives OBS: one 4K camera, three angles, audio-driven switching with voiceprint support for shared microphones. Free while in public validation.

Download AutoDirector →

Raw logs, scoring scripts and the exact test material are published. Reproduce the numbers before you believe any vendor's table — including ours.