Voice Print Matcher — same speaker?
By Ned Walsch · Last updated:
Load two audio clips and get a same-speaker similarity score — the audio equivalent of matching a face. Compare an anonymous threatening call against a known sample, or check whether two voice notes are the same person. A speaker-embedding model runs entirely in your browser; the audio is never uploaded.
A score is a lead, not a verdict. This measures how consistent two voices are, not whether they are legally the same person. Forensic voice comparison is genuinely difficult: recording quality, noise, illness, emotion, and deliberate disguise all move the number. Treat a high score as a reason to investigate further and a low score as a reason for caution — never as courtroom-grade proof, and never as a substitute for a qualified examiner.
Where the processing happens
Each clip is turned into a speaker embedding — a numeric fingerprint of the voice — by a WavLM model fine-tuned for speaker verification, running on WebAssembly with WebGPU acceleration where available. The tool then compares the two fingerprints with cosine similarity. The model is fetched once from a public CDN and cached; the audio never leaves your device, which you can confirm in your browser’s network tab.
How to read the score
The score runs from 0 to 100. Higher means the two voices are more consistent with being the same speaker. As a rough guide, very high scores suggest a likely match, mid-range scores are genuinely uncertain and should not be leaned on, and low scores suggest different speakers. The exact boundaries depend on clip quality, so use the score to rank and prioritize, not to decide.
What breaks it
Short clips (under a few seconds), background noise, music, overlapping speakers, phone-line and codec compression, and big differences in recording conditions between the two clips all reduce reliability. A person can also sound different when shouting, whispering, ill, or intoxicated. If two clips were recorded on very different devices, expect the score to drop even for the same speaker.
Responsible use
Voice comparison touches privacy and, in some places, biometric-data law. Only analyze recordings you have the right to use, be cautious about drawing conclusions that could harm someone, and remember this tool cannot and does not make an identification. For a matter that carries real consequences, consult a qualified forensic voice examiner. See Is OSINT legal?
Can this prove two recordings are the same person?
No. It gives a similarity score that is a strong lead, not proof. Recording quality, noise, emotion, illness, and disguise all affect it. A high score means the voices are consistent with the same speaker; it is not a courtroom conclusion and not a replacement for a forensic examiner.
How does it work?
Each clip is run through a WavLM speaker-verification model that turns a voice into a fingerprint capturing vocal characteristics rather than words, then the tool measures cosine similarity between the two fingerprints. Everything runs locally; the audio is never uploaded.
What clips give the best result?
Clean recordings of at least a few seconds of continuous single-speaker speech, with little noise and no overlapping voices. Trim multi-speaker clips to the person you care about first.
Is my audio uploaded?
No. The model is downloaded once and cached; all analysis happens in your browser. The audio never leaves your device.