Can you separate two singers or multiple voices in one recording?
Sometimes, but the recording and the meaning of “separate” matter. A vocal/music model separates voice-like material from accompaniment. A lead/backing model separates musical roles. Neither is automatically a system that recognises two people and gives each an independent file. If two singers share the lead melody, a lead/backing result may keep them in the same group.
1. One-minute audio test
Slow Fade Duet with Mr_Yesterday — Speck, featuring Mr_Yesterday; lyrics D A Ayer · 1:38–2:38 · 60s
Lead/backing separation does not identify individual singers. Compare the overlapping voices rather than treating the outputs as singer 1 and singer 2.
Click or drag a waveform to seek. On a focused waveform: ←/↓ back, →/↑ forward; Tab next track, Shift+Tab previous track. Space plays or pauses.
Loop
Test settings and processing record
Tested on 7 November 2026: standard mode, 4-second segments, NVIDIA RTX 4060. Times cover local inference and export, excluding model loading, queue and transfers.
| Model | Processing time | Peak VRAM |
|---|---|---|
| gabox-karaoke-v2 | 7.56 s | 1,449.2 MiB |
Model loading, queue and transfer time are excluded. All versions use the same input and standard mode. MP3 previews: 192 kbps; references: 96 kbps mono, 16 kHz.
Author reference audio — not AI output
Reference recordings were aligned to the original mix to check that the target sound occurs in this passage. This verifies activity, not subjective separation quality. Reference levels were adjusted; these are not the original mix balance.
Full test record ↓2. Identify the overlap, then choose a separation method
Listen to the difficult moment and answer three questions. Are the two people singing or speaking? Do they alternate, or sound at the same time? Are their recordings already on different channels or available as separate source files? These answers determine the next step more usefully than the number of people visible in a video.
Online models cover vocals/instrumental, four or six stems, instrument extraction, lead/backing vocals and dereverb. Check availability in the selector. Outputs are estimates, not named-speaker identification, general noise repair or guaranteed recovery of every sound.
- If you need both singers without instruments, start with ordinary vocal isolation.
- If one part is a clear melody and the other a quieter harmony, test a lead/backing option on the overlap.
- If both voices carry equal melodic focus, do not assume the louder or higher-pitched singer will always become Lead.
- If a singer changes role later, check that section separately rather than trusting the first successful phrase.
After the full result is ready, compare the difficult passage with the original. Check both the wanted detail and unwanted leakage before choosing a track.
3. Check the target voice and the limits of separation
Pick a phrase where both voices overlap. Follow the wanted voice through the first consonant, sustained note and ending. Next, follow the unwanted voice and ask whether it is still intelligible. Finally, return to an alternating phrase and check whether the wanted person remains in the same output. This is a suggested audition routine, not an identity-recognition test that the website has passed.
If the wanted voice disappears whenever its role changes, the result is unsuitable for a continuous one-person edit. Do not repair that conclusion by describing the output as “nearly isolated” without checking the intended use. A learning aid can tolerate more spill than an exposed vocal edit, but both need an honest label.
A person may be placed near the center, but so may another singer and several instruments. Centre-based cancellation relies on stereo relationships, not identity, and wide vocal effects can survive it. It is not a reliable substitute for a dedicated person-separation system.
4. For speech, distinguish labels from audio
Knowing who spoke at a timestamp is not the same as generating a clean waveform for each person. A transcript with speaker labels can help navigate a recording while simultaneous speech remains mixed. uvr online’s current music modes should not be sold as a speaker-diarization or overlap-repair service. For an interview you control, ask for the independent microphone recordings first; for serious overlap, evaluate an appropriate speech-editing tool and verify the entire segment.
If the result cannot preserve the requested person, keep the original and explain the limitation. Authorized source tracks, separate microphone feeds or a new recording can be a better answer than promising that another music model will separate any two people.
5. Frequently asked questions
Can I remove one particular person’s voice?
Not as a general supported promise in the current music workflow. Role-based models may separate a backing part, but they do not let you identify any person and erase that person everywhere.
Does lead vocal always mean the first singer?
No. Lead describes the model’s estimate of musical focus. It is not a stable singer number or identity label.
Will speaker labels make separate audio files?
No. Labels indicate who speaks when. Independent audio separation is another operation, particularly when voices overlap.
Can I separate voices in audio if they never overlap?
Manual time selection may be enough to collect alternating phrases. It cannot create a continuous independent performance during any overlap or reconstruct words that were never recorded on their own.
Audio sources and usage
Slow Fade Duet with Mr_Yesterday — Speck, featuring Mr_Yesterday; lyrics D A Ayer · CC BY 3.0
Includes source material from: Mr_Yesterday / 59439
The excerpt was trimmed, separated into audio stems and encoded as MP3.
If a sample affects your privacy or rights, contact support@uvr.online or open a Support ticket with the sample link to request a review and removal. Contact support →