Speech separation

Create two estimated speech sources as new mono clips.

Try it with my recording Find help for my audio problem

Audio editor showing a real recording waveform, spectrogram and tools — A real recording in the editor. This screenshot demonstrates the workflow; evaluate treatment results using your own audio.

When should you choose this tool?

Separate versions of two overlapping voices can help examine individual utterances. This tool creates two outputs while retaining the source. Always assess the estimates alongside their common original.

How it changes audio

Neural separation estimates two speech sources from 16 kHz mono input. New clips return to the project sample rate, but that does not restore removed high frequencies.

What do the parameters control?

Selection: Choose the passage on the timeline. Shorter passages require fewer resources and are easier to review.
Preview and two outputs: There is no sensitivity or speaker-count slider. Preview plays the first estimate; Apply creates two new mono clips.
Initial download: The editor checks availability and displays the initial model download size. Downloading and local computation can take time.

Try a short passage

Begin with a short two-speaker passage. Listen to the original so you know which utterance you are examining.
Wait for the availability check, then start Preview. Apply if the estimate is useful.
Listen through both new clips and the original. After pauses, check whether the output roles have swapped.

What to listen for

One output may clarify a phrase while the other is more useful elsewhere. Do not assign permanent speaker identities based on the opening seconds.

Limits of the method

Outputs can swap order after silence. This is not transcription, authenticated identity or guaranteed isolated tracks.

Preview covers a short passage. After applying, review the full treated section; you can undo the step from history.

Other useful tools

Speech cleanupSpeech-focused noise reduction with adjustable blend and multiple processing modes.Extract speakerFollow a chosen voice using a clean reference passage from that speaker.Spectral fusionCompare and spectrally combine treatment candidates made from the same input.