In this guide
Check for separate tracks or microphone channels first. When people take turns, a manual edit can keep each turn. When voices overlap in the same channel, test source isolation and listen for missing words; a clean track for each person is not guaranteed.
To separate two voices in one audio recording, first find out where the voices are stored. Two microphones might already have separate files or left and right channels. Two people on one microphone may share a single mixture. These cases need different edits.
Start with the original recording and a short passage where both people can be heard. If the speakers overlap, include that overlap in your test. A result that handles their alternating turns does not establish that it can recover two simultaneous sentences.
Choose a workflow from the recording you have
| Your source | Start here | Main check |
|---|---|---|
| A separate recording for each microphone or participant | Import the original tracks into an editor | Each file may still contain spill, room sound, or a timing offset |
| One stereo file with a different microphone on each channel | Extract the existing left and right channels | Listen to both sides; stereo does not guarantee one clean voice per side |
| A mixed file where speakers take turns | Copy the timeline and silence the other person's turns | Preserve pauses, breaths, and the original start position |
| Two voices speaking together in the same channel | Test source isolation on the actual overlap | Check whether the wanted words survive and the other voice remains audible |
If you can obtain the separate microphone recordings, use them as your first editing candidates. Do not process a finished mix simply because a separator is available. The task is to get a useful voice track with as little unwanted change as practical.
Check left and right before using AI
A stereo file can contain two microphone feeds, a complete mix on both sides, or a voice on one side and an unrelated recording on the other. A waveform or the word “stereo” does not tell you which arrangement you have.
- Open a supported short stereo MP3 or WAV in the free audio channel splitter.
- Prepare Left channel and listen to the passage where both people speak.
- Prepare Right channel and listen to the same passage.
- If a side already has the voice you need, download that mono WAV. Save the other side too if it contains useful conversation or ambience.
Channel extraction keeps the sounds recorded on that side. It does not remove the second speaker if their voice was picked up by the same microphone. Averaging both channels into mono combines them; it does not identify a person.
For a desktop workflow, Audacity documents splitting stereo into separate tracks or mono tracks. Center a single voice track deliberately in your edit instead of leaving a wanted interview voice on only one playback side. The channels and stems guide explains the distinction in more detail.
Saturalabs' free splitter accepts stereo MP3/WAV up to 50 MiB and five minutes, and exports 48 kHz, 16-bit mono PCM WAV locally. Its free downloads do not use the separate AI service's account or credits.
If people take turns, edit turns without closing the gaps
When only one person is speaking at a time, an editor can keep that person's turns and silence the other turns on a working copy. This is a timing edit of the mixture. Background noise and anything else audible during the retained turns remain.
Duplicate the source onto two tracks with the same start position. On the first copy, silence the second person's non-overlapping turns. Reverse that choice on the second copy. Leave the silenced intervals in place if the files need to line up with each other or with picture. Deleting an interval and moving later speech earlier changes that alignment.
For an illustrative 12-second clip, suppose person A speaks at 0–3 seconds, person B at 3–7, and person A again at 7–12. A's timeline-preserving track keeps 0–3 and 7–12, with silence from 3–7. It still lasts 12 seconds. Joining only A's turns into a new 8-second file makes the second turn begin at second 3 rather than second 7.
Mark boundaries by listening. Keep word endings and useful breaths, and use short fades where an abrupt edit creates a click. If both people are speaking from 6.5–7 seconds, keeping that interval retains both voices; silencing it loses both. A turn edit cannot resolve that overlap.
Test overlapping voices with a focused source description
For mixed speech that cannot be recovered by choosing a channel or editing turns, the voice isolator lets you try one audible source description. Use a representative original excerpt, not a previously damaged separation. Upload a supported MP3, WAV, MP4, or MOV up to 50 MiB (52.43 MB).
A generic “spoken voice” request can include both speakers. Try a characteristic you can actually hear that distinguishes the wanted voice. The following are candidate descriptions to evaluate, not tested commands that guarantee a particular speaker.
| What distinguishes the wanted voice | A description to try | What to listen for |
|---|---|---|
| A noticeably lower pitch | The lower-pitched person speaking | Whether the higher-pitched voice stays attached during overlap |
| A foreground narrator against distant conversation | The foreground narrator's spoken voice | Soft narrator syllables and background words between sentences |
| A breathy voice against a clearly different tone | The breathy speaking voice | Whether breaths and quieter words survive with the voice |
A person's name, job title, or the label “speaker 1” does not explain how they sound. Do not assume the service maps transcript labels or real-world identities to voices. Use the prompt guide to change one audible detail at a time and compare the same excerpt.
Inspect Isolated sound for the requested voice and Background audio for the remaining mix. One pass returns a target and a remainder. The remainder is not automatically a clean second-speaker track: it can include music, room sound, other people, and fragments of the first voice.
If you need to test the second voice separately, make a new request from the same original with its own description. Independently estimated outputs are not guaranteed to add back to the original mix. Downloads require an account and sufficient credits; the result screen shows the cost, and pricing explains account options.
Review five sections before choosing a voice track
Compare the original and candidate at a similar listening level. Loudness alone can make a result seem more successful. Use the same five sections for each attempt, and record the source time so you can return to them.
| Section to mark | Question to answer | Reason to reconsider |
|---|---|---|
| Wanted speaker alone | Are the entire sentence and its quiet ending intact? | Missing consonants, changed tone, or a word cut short |
| Other speaker alone | Is their unwanted speech still intelligible in the candidate? | The result may be general speech isolation rather than one-person isolation |
| Both speakers together | Can you follow the wanted sentence through the overlap? | Missing wanted words or both sentences staying together |
| A handover, breath, or brief interruption | Does the voice remain continuous through the change? | A dropout, unnatural cut, or distracting texture |
| The final edit | Does the candidate remain useful at the intended placement and level? | Timing drift, doubled speech, or silence that breaks the scene |
A useful note is specific: “A's last word survives at source 00:42.8, but B is still intelligible at 00:43.1.” Keep the prompt and output choice beside that note. It helps distinguish unwanted spill from damaged wanted speech and makes another test easier to compare.
These checks are an editing workflow, not measured accuracy results. Similar voices, sustained overlap, strong echo, or clipping can limit a candidate. If the wanted sentence is damaged, retain the original passage, use a separately recorded take when available, or choose an edit that does not require clean isolation of that overlap.
Speaker diarization is different from separate audio tracks
A transcript that labels turns as “speaker 1” and “speaker 2” tells you who spoke when. It does not by itself give you a clean WAV for each person. Google Cloud's speaker diarization documentation describes numbered speaker labels on transcribed words.
Those labels can help you mark turns for an edit. During simultaneous speech, copying a labeled time range still copies the whole mixture in that range. Saturalabs' isolation workflow separates a described audio source; it does not produce a speaker-labeled transcript or automatically export one track for every detected participant.
Keep the timing when returning voices to a podcast or video
Give outputs clear names that include the source range, target description, and chosen track. Keep independent tests of both people based on the same excerpt so their intended start position is traceable. Do not remove quiet intervals simply to make a file look more active.
If an excerpt covers source 42.50–68.00 seconds and the full source starts at timeline zero, place its output at timeline 42.50. For a clip already trimmed or moved, calculate the corresponding position in your editor instead. The audio replacement and sync guide gives the full offset calculation and final-export checks.
Check a word near the beginning and one near the end. When auditioning a replacement, mute the corresponding original audio so you do not hear the voice twice. If using independently separated tracks together, listen for duplicated fragments and tonal changes rather than assuming their sum is a faithful reconstruction.
For the next recording, keep individual microphone or participant recordings when your recording setup provides them. Review those files before the session ends. Separate capture gives you editing access that a final mixed file may no longer provide.
Common questions
Can I separate two voices from a mono recording?
You can edit non-overlapping turns manually or test AI isolation for a distinguishable voice. A mono file has no hidden left and right channels to extract. Clean recovery of both voices during overlap is not guaranteed.
Can I remove the interviewer but keep the guest?
Check for separate tracks or channels first. For a mixed file, keep the guest's non-overlapping turns with a timing edit, or test a description of the guest's audible voice. Check interruptions and every wanted word that overlaps a question.
Will a vocal remover split two speaking people?
A singing-versus-instrumental task does not specify which speaking person you want. Start with a speaking-voice workflow and an audible distinguishing detail. A general voice extraction can still retain both people.
Is separating left and right free?
The channel splitter provides local channel extraction and WAV downloads without an account or credits for supported short stereo files. AI source isolation is a different workflow with account and download-credit requirements.