Audio Editing

How to separate two voices from one audio?

Separating two voices depends on how they overlap. Check for separate channels, split turns by hand, or try AI speech separation for overlapping talk.

On this page
  1. Short answer
  2. Check whether the voices are already separate
  3. Speakers who take turns
  4. Voices talking over each other
  5. Tools that don’t do this
  6. Next time: record each person separately

Short answer

The method depends on how the two people were recorded.

  • If each person is on a separate channel, split the channels.
  • If they take turns, cut each person’s sections onto their own track.
  • If they talk over each other, AI speech separation can help, but expect artifacts.

Recording each person on their own microphone and track is the most reliable fix, and it’s worth doing next time.

Check whether the voices are already separate

Open the file and look at its channels.

  • Stereo file with one person per side: In Audacity, open the track’s drop-down menu and choose Split Stereo to Mono. That gives you two tracks.
  • Mono file: There are no separate channels to split. Use one of the methods below.

Audacity’s Vocal Reduction and Isolation effect (Effect > Special) separates centre-panned sound from side content. It only helps when the voices sit at different positions in a stereo mix. It can’t separate two voices in a mono file, or two voices that are both centred.

Speakers who take turns

Splitting by hand is slow, but it’s reliable.

  1. Select one speaker’s section.
  2. Press Ctrl+Alt+I (Split New). This moves the selection onto a new track.
  3. Repeat for each section, then line the clips up. For splitting and moving clips, see audio editing basics.

Diarization tools can speed up the labelling. They detect who is speaking and when, so you can jump to each section. Accuracy varies, so listen to the results. Examples include pyannote.audio, an open-source toolkit, and some transcription editors that add speaker labels.

Voices talking over each other

This is the hard case.

  • EQ alone can’t separate two voices, because their pitch ranges overlap. As a rough guide, adult male speaking fundamentals sit around 85 to 180 Hz and adult female ones around 165 to 255 Hz. Harmonics spread across the same frequencies.
  • Noise reduction targets steady noise, not a second voice.
  • Speech separation models try to split overlapping speech into separate tracks. Open-source options include SpeechBrain’s SepFormer and the Asteroid toolkit’s Conv-TasNet. They’re often trained on two or three speakers at 8 or 16 kHz, and results vary with the recording.

Expect some artifacts, bleed from one voice into the other, and a change in tone. Keep the original file, and listen to each output before you use it.

Tools that don’t do this

  • Music stem splitters, such as Demucs, Spleeter, and UVR, separate vocals from instruments. They don’t separate one speaker from another.
  • Speech cleanup tools, such as iZotope RX’s Dialogue Isolate, separate speech from background noise. They don’t separate speakers.
  • Audacity’s AI-based effects in recent versions cover music separation and noise suppression, not speaker separation.

Next time: record each person separately

The most reliable way to separate two voices is to avoid mixing them in the first place. Put a microphone close to each speaker, and record each one on its own track or channel. On a call, Zoom can save a separate audio file for each participant.

For cleaning up the separated tracks, see improving audio quality in Audacity, which covers noise reduction and EQ.

The complete guide to audio editing

Audio Editing Basics: Splitting and Exporting Tracks in Audacity
Read the guide

Keep exploring

Guides from other topics

All topics →