Pick by job
Which tool for which job
Quick answer: for a person talking over a music bed, Diffio removes the music and cleans the voice in one pass, best when the room is also noisy or echoey. For songs, use a vocal separator: Diffio is not built for singing.
A song: you want the vocals or the instrumental
A vocal separator: BS-RoFormer, UVR, Demucs, LALAL.AI or remove.music
Built for sung vocals; on our 6 songs they kept the lyrics far better than Diffio. BS-RoFormer, UVR and Demucs are free on your computer; LALAL.AI downloads cost $9.99 a month.
Dialogue, music and effects as three tracks
Audo Remove music, or Adobe Podcast's paid plan
They return the music apart from the effects; Diffio returns two tracks. Audo is 8 credits ($0.08) a minute, with the first 45 seconds free 3 times a day.
The voice waveform kept exactly as recorded
A vocal separator
Diffio rebuilds the voice; separators keep it far closer to the original recording.
Someone talking over music, in a noisy or echoey room
Diffio
Music, noise and echo removed in one pass. On our 8 noisy or echoey clips with music: clearest speech and lowest word error.
Someone talking over music, and it should sound finished
Diffio
Highest naturalness on 35 of 36 clips; no music detected in the speech on any of them.
Listen
Hear the test clips, the failure too
Dashed outline: a simulated mix, not a real recording.
A sermon over an electro-pop trackSimulated mix
A Sunday sermon with an electro-pop bed 8 dB under the voice.
Listen for: The typical case: music gone, voice clearer. Naturalness 2.24 → 4.48 (separator 3.35); word errors 1.6% → 0%.
Source: Speech: "Sunday Sermon | The Gospel of Mark: What is Jesus Worth?" by FBC Boerne (YouTube, CC BY 3.0). Music: "Cloud Dancer" Kevin MacLeod (incompetech.com), CC BY 4.0.
Source & method
License: CC BY 3.0 and CC BY 4.0. Source recording. Full credits. A test mix made from a separate voice and music recording, so the clean voice is known. Run through Diffio 4.7 Pro and BS-RoFormer (a free vocal separator, default settings) on November 19, 2026. Every version is level-matched to the same loudness; a silent result stays silent.
- ALMA (orchestra)2.76 → 4.79
- Old Glory (full band)2.82 → 4.84
- Air Force history (marching score)2.28 → 4.85
- Wanna Work Together? (upbeat)1.68 → 4.71
- GLAM-Wiki (interview)3.38 → 4.66
Measured
55 clips, every result kept
36 people talking over music, 8 noisy or echoey recordings with music added, 6 songs and 5 real videos. For each we know which part is speech and which is music.
- Music gone
- 36 of 36
- talk-over-music clips with no more music detected in Diffio's speech than in the clean voice
- Sounds finished
- 4.52 / 5
- naturalness of Diffio's speech; the mix 2.68, the clean voice 3.82
- Noisy room + music
- 4.63 / 5
- on 8 clips; the free separators reached 2.73 to 3.11
- Singing
- 1 of 6 silent
- Diffio is not a vocal separator; on the other 5 songs it kept a quiet, rebuilt vocal
By kind of video: naturalness and word errors
| Speech over music | Clips | Naturalness, mix → Diffio | Word errors, mean | Word errors, median |
|---|---|---|---|---|
| Vlog | 6 | 2.68 → 4.52 | 24.2% → 15.4% | 7.8% → 8.8% |
| Interview / podcast | 6 | 2.61 → 4.41 | 7.1% → 6.3% | 6.8% → 4.4% |
| Lecture | 6 | 2.84 → 4.61 | 14.6% → 4.0% | 2.8% → 5.1% |
| Sermon | 6 | 2.77 → 4.62 | 2.8% → 2.3% | 2.3% → 0.9% |
| Ad read | 6 | 2.66 → 4.52 | 6.5% → 6.0% | 3.7% → 1.7% |
| Gameplay commentary | 6 | 2.49 → 4.41 | 6.8% → 1.9% | 4.0% → 2.2% |
| Noisy or echoey room + music | 8 | 2.27 → 4.63 | 8.5% → 5.5% | 5.1% → 3.2% |
Word error compares a machine transcript with a reference; differences under about 3 points are noise in this set. Vlogs score high for every tool, the clean voice included (14.5%), because one clip is half a street conversation in broken English.
Limits
What Diffio doesn't do
It rebuilds the voice; it doesn't cut it out. Against the clean voice, BS-RoFormer matches the waveform at 23.0 dB SI-SDR, Diffio at 1.4 dB; voice similarity 0.97 vs 0.86. If you need the original waveform, use a separator.
The music track is better from a separator. To keep the music, BS-RoFormer's music track matched the real music at 16.2 dB; Diffio's at 8.8 dB.
Singing isn't separated. On 6 songs Diffio's speech track got 61.6% of the lyrics wrong (separators 22 to 24%, the clean vocal 19.4%), and on one song it came back silent.
In noisy rooms the voice changes more. On the 8 noisy or echoey clips, similarity to the original recording was 0.73 (separators 0.92 to 0.95), as low as 0.47 on a phone-line recording, because Diffio also removes the noise and echo the original keeps.
Word error on clean speech under music is about the same for every tool: mean 6.0% Diffio, 7.1–7.3% separators; Diffio was higher than the best separator on 15 of 36 clips.
Show the full table: Diffio and three free separators on the same 36 clips
| 36 clips (means) | Mix | BS-RoFormer | UVR-MDX Voc_FT | Demucs htdemucs | Diffio 4.7 Pro | Clean voice |
|---|---|---|---|---|---|---|
| Naturalness, SCOREQ 1–5 | 2.68 | 3.88 | 3.84 | 3.69 | 4.52 | 3.82 |
| Word error, mean / median | 10.3% / 3.9% | 7.2% / 3.6% | 7.1% / 3.6% | 7.3% / 3.3% | 6.0% / 2.7% | 7.7% / 3.5% |
| Music detected (AudioSet), highest clip | 0.61 | 0.029 | 0.152 | 0.079 | 0.017 | 0.041 |
| Clips with more music detected than the clean voice | 36 | 0 | 6 | 6 | 0 | – |
| Speech track in pauses vs the mix (median, 16 clips with pauses) | 0 dB | −37 dB | −30 dB | −26 dB | −36 dB | −36 dB |
| SI-SDR vs clean voice (waveform match) | 11.7 dB | 23.0 dB | 20.8 dB | 19.9 dB | 1.4 dB | – |
| Voice similarity to clean voice | 0.95 | 0.97 | 0.96 | 0.97 | 0.86 | 1.00 |
| Music track vs real music (mel SDR) | – | 16.2 dB | 3.8 dB | 11.6 dB | 8.8 dB | – |
The free separators ran with default settings (BS-RoFormer at overlap 2, not its default 4, for time). Diffio 4.7 Pro ran on November 19, 2026. Audo's Remove music (free, 3 clips) showed the same picture as the separators: naturalness 3.68 vs Diffio 4.44, voice similarity 0.94 vs 0.82, no music detected in either.
On the one narration-over-music clip in our speech enhancer test, speech tools that don't split stems came close: naturalness 3.00 for the original, 4.73 Adobe Podcast, 4.68 Diffio 4.7 Pro, 4.59 Auphonic, 4.46 Cleanvoice, 4.44 ElevenLabs. Adobe scored higher than Diffio, and Diffio got more words wrong (6.3%) than the original or ElevenLabs (1.3%). One clip only.
Weather explainer over music (speech enhancer test)
Source & method
A narrated explainer with a music bed, run through every tool in our speech enhancer test. Source: "El Niño – WMO animated explainer video", World Meteorological Organization, CC BY 3.0, via Wikimedia Commons.
License: CC BY 3.0. Source recording. Restored with Diffio 4.7 Pro on November 3, 2026; Adobe Podcast and ElevenLabs were run on the same clip with the settings listed in the test, for our 59-clip speech enhancer test. Every version plays at matched loudness.
How to
Remove music from a video in 3 steps
- 01
Upload your video
MP4, MOV, AVI, MKV or WebM, up to 90 minutes and 2 GB.
- 02
Listen with the music off
Drag Background amount to 0% for speech only, or leave a little music under the voice.
- 03
Download the video
MP4 keeps the original picture. Or download the speech, the music or the mix as WAV, MP3 or M4A for your editor.
Free shows a 2-minute preview of each result; downloads need Basic or Pro.
Plans
Free to try, then by the hour of video
Free
$0
Try Diffio on your own recordings.
- 3 uploads in total (not monthly), audio or video
- Preview the first 2 minutes of each result
- Diffio 4.7 Flash model
- No full playback, download or export
Basic
$8per month
For hobbyists and small teams.
- 6 hours (360 minutes) of audio or video a month
- Full playback, download and export
- Diffio 4.7 Flash and 4.7 Pro. Developer model docs
Pro
$29per month
For professionals and businesses.
- 30 hours (1,800 minutes) of audio or video a month
- Full playback, download and export
- Diffio 4.7 Flash and 4.7 Pro. Developer model docs
Diffio costs $1.60 an hour on Basic at 5 hours a month ($8 a month for up to 6 hours). A video counts by its length.
Questions
Questions about removing music from video
Can I remove background music from a video and keep the voice?
Yes, when the voice is speech over a music bed. Diffio separates the speech from the music and returns the same video with the music removed or turned down, plus the speech and the music as separate tracks.
Do I get a video back or just audio?
Both. Video projects download as MP4 with the original picture; you can also download WAV, MP3 or M4A audio of the speech, the music or a mix.
Can it remove music from a song and keep the singing?
No. Diffio is built for speech; on our 6 test songs it lost most of the lyrics and on one it returned silence. Use a vocal separator such as BS-RoFormer, UVR, Demucs, LALAL.AI or remove.music for songs.
Does it remove noise and echo too?
Yes, in the same pass. On 8 noisy or echoey recordings with music added, the speech scored 4.63 out of 5 for naturalness with Diffio, against 2.73 to 3.11 for free separators that remove the music only. See removing background noise from video.
Will the voice sound different?
A little. Diffio rebuilds the voice rather than cutting it out, so it measures 0.86 similar to the clean recording on average, where separators measure 0.96 to 0.97. Keep the original and compare.
4 more questions
Is it free?
The Free plan gives 3 uploads with a 2-minute preview of each result, and no downloads. Basic is $8 a month for 6 hours, which works out to $1.60 an hour at 5 hours a month. See plans.
How long can the video be?
Up to 90 minutes and 2 GB per file in the web app. Through the API there is no length limit, up to 2 GB per file.
Can I keep the music and remove the voice instead?
Yes, the music is a separate track you can download, but a vocal separator gives a cleaner music track (16.2 dB vs 8.8 dB against the real music in our test).
Can I get dialogue, music and effects as three tracks?
Not from Diffio, which returns two (speech and background). Audo's Remove music and Adobe Podcast's paid plan return separate music stems. For a voice and background split of any recording, see the voice isolator.