Speaker similarity
Compare learned speaker embeddings from two voice recordings, locally.
ECAPA-TDNN speaker embeddings + cosine similarity. First model download: 42 MB. Uncalibrated score, not identity proof or probability.
Loading your audio workspace…
How to use this tool
- 01
Choose a recording
- 02
Load the model and run locally
- 03
Review the result, then export
New processing and analysis tools accept browser-decodable audio up to 20 MB and 60 seconds per file, with at most two channels. Longer files can be cut into shorter clips. Processed audio exports as 16-bit PCM WAV.
Does my recording leave my device?
The available upload-processing tools decode and process audio in this browser. Microphone tools ask for permission. System text-to-speech may use an online voice provider when an online voice is selected. Available AI tools download models after you start, then process audio locally. Planned tools do not accept files.
What should I know about the results?
ECAPA-TDNN speaker embeddings + cosine similarity. First model download: 42 MB. Uncalibrated score, not identity proof or probability.
Try a short clip first and listen before sharing. Processing can introduce artifacts. Analysis describes the recording, not the identity, health or trustworthiness of a person.
Keep working on your voice
New tool descriptions and controls support English and 简体中文. Navigation follows your selected interface language.