Separate overlapping voices in a conversation recording
Two estimated speech sources or reference-guided extraction: which to try and how to review it.
Extract speaker
Follow a chosen voice using a clean reference passage from that speaker.
When should you choose this tool?
To follow one person in a conversation, first find a clear passage where that person speaks alone. The reference guides processing toward the chosen voice. Without one, two-source Speech separation may offer an alternative to compare.
How it changes audio
Reference-guided neural estimation attempts to extract the target voice. It processes 16 kHz mono audio and then returns to the project’s sample rate and channels. It is not speaker identification.
What do the parameters control?
- Speaker reference; New reference name; Learn
- Select one clean 3–6 second speech passage and name it. Up to 3 references are supported. Its name is your label, not a verified identity.
- Extraction strength
- A 0–100% blend. Gentler settings retain more original sound; stronger settings cannot compensate for a poor reference.
- Delete reference
- A reference used by an earlier step cannot be removed while replayable processing needs it.
Try a short passage
- Find a clear solo utterance from the target voice. Avoid overlapping speech in the reference.
- Name and learn the reference, then select the conversation passage and start Preview.
- Check quiet words and overlaps separately. If words disappear, try a different reference or gentler blend.
What to listen for
The target voice should remain trackable when another voice enters. Target speech in Residual may reveal lost content.
Limits of the method
Stable separation throughout a recording is not guaranteed. Mono speech processing cannot preserve original high frequencies and stereo space.
Preview covers a short passage. After applying, review the full treated section; you can undo the step from history.