MotionShine01 / Words
Written out
Whisper, an open speech model, writes out what is said, with the time of each line.
Free tool
Who said what, in your transcript.
Drop in an interview, a podcast, a meeting or a call, and get a transcript that marks each speaker: Speaker 1, Speaker 2 and so on, with the time each one starts. Rename them to real names, then copy the text or save it as subtitles. Speech recognition and speaker detection run in this tab, so the recording is never uploaded.
Recording
Names
Speakers appear here to rename once they are found.
Your recording is transcribed in your browser. Nothing is uploaded.
How it works
Words, then voices.
02 / Voices
Told apart
A second model hears who is talking, and a voice print matches the same person all the way through.
03 / Names
Rename and save
Change Speaker 1 to a name, then copy it or save text, SRT or VTT with the names in.
Questions
Good to know.
Is my recording uploaded?
No. The models are downloaded to your browser once and run there. Interviews and calls often hold private details, which is why nothing is sent to a server.
Which models does it use?
Whisper base (OpenAI, Apache 2.0), in a build by onnx-community that times every word, writes out the speech. pyannote segmentation 3.0 (MIT licence) hears when each voice talks, and the WeSpeaker ResNet34 voice print model (CC BY 4.0, by the WeSpeaker team) matches the same person across the recording.
How accurate is it?
It works best with clear recordings where people take turns. Short replies like "yes" may go to the wrong person, and voices that sound alike can be merged. If you know how many people talk, set the number: it helps.
Can it put real names on the speakers?
It cannot know names, but you can type them in the Names list and every line updates.
What about a recording without speakers?
For one voice, audio to text is quicker and gives the same transcript with subtitles. For notes and actions from a meeting, use meeting notes.