Can You Compare Two WhatsApp Voice Messages to See If It's the Same Person?
Two voice messages arrive from two different numbers and you want to know whether the same person sent both. It is one of the most common reasons people use a voice comparison, and voice notes are close to the hardest input such a tool gets: short, compressed by the app, recorded on a phone at arm’s length, often with a room or a street behind them.
Here is what that does to the answer, measured rather than guessed.
What the app does to the recording
Messaging apps encode voice notes with Opus, a codec built for speech at very low bitrates — often 12 to 24 kbps. It keeps speech intelligible and throws away a lot of everything else. Add room echo and background noise and there is less of the voice left to measure.
What we measured
We took 200 pairs of recordings where both sides are genuinely the same person (VoxCeleb, a public research set), encoded one or both sides the way a messaging app sends a voice note, and ran them through exactly the engine the voice comparison uses.
| what happened to the recording | called “same” | “not conclusive” | called “different” |
|---|---|---|---|
| Opus 24 kbps, one side | 83% | 17% | 0% |
| Opus 12 kbps, one side | 75.5% | 24.5% | 0% |
| Opus 12 kbps, both sides | 79% | 21% | 0% |
| Opus 16 kbps + room echo | 73.5% | 26.5% | 0% |
| Opus 16 kbps + background noise | 59.5% | 40% | 0.5% |
The pattern is the useful part. Compression, echo and noise make “Not conclusive” more likely, and almost never turn the same person into a confident “different”. When the evidence gets thin, the tool says it does not know rather than guessing.
The thing that actually goes wrong: length
When people told us a voice comparison got it wrong on voice messages, the codec was not what those cases had in common. Length was.
The free plan analyses the first 15 seconds of speech in each recording. If a message runs a minute and a half and someone else talks first — a forwarded clip, a second person in the room, the other side of a call — those first 15 seconds may not be the voice you mean. The result says when a recording had more speech than was read, and which one.
How to give it the best chance
- Ten seconds of speech or more on each side. Below that is where most of the undecided results happen. Five seconds of “hi, call me back” is not much to go on.
- Trim to the voice you mean. If the person you care about is not the first to speak, cut the recording so they are.
- One speaker per file. A conversation hands the model two voices mixed together.
- Upload the original file. Save the voice note itself (
.opusor.ogg) rather than playing it through a speaker and recording it again.
What the answer can and cannot tell you
A “same speaker” verdict means the two voiceprints sit where recordings of one person usually sit. It is evidence, not proof of identity, and nothing a voice model produces should be treated as proof on its own.
It also cannot see through a voice-changer app, which moves exactly the characteristics the model listens to. If you suspect one, read can voice comparison detect a disguised voice before reading much into a “different”.
Running the same pair again gives the same answer. If the result was “Not conclusive”, the way forward is longer or cleaner recordings, not another try.
Stop hunting for differences by hand. DiffALL spots every change between any two files — automatically.
Compare your files — free