How can you use AI to slow down the BPM of a song?
Slowing tempo without dropping pitch is time-stretching, and it's decades-old DSP, not AI. Where AI genuinely helps is stem separation: models like Demucs (MIT license, built at Meta) split a track into drums, bass, vocals, and other, so each part stretches cleanly instead of smearing into one muddy chipmunk.
Why — the first-principles explanation
Sound is pressure waves in time. Play a recording at 80% speed and everything gets 25% longer and 25% lower — speed and pitch are welded together, which is why slowed-down tape sounds like a monster. Separating them is the whole problem.
The classic fix is the phase vocoder. Chop the audio into overlapping slices, convert each to its frequency content, then re-lay the slices further apart in time while keeping their frequencies untouched. Longer duration, same pitch. This is ordinary signal processing and it's been in Audacity, Ableton, and every DJ deck for years. No AI required, and none used.
So why does it sound bad? Because the trick assumes each slice contains one steady thing. Real music doesn't. A drum hit is a sharp transient — a click in time — while a vocal is a sustained tone. Stretch them with the same setting and you either smear the drums into a flam or wobble the vocal into that watery, underwater artifact. The algorithm can't satisfy both, because it can't tell them apart. That's the actual bottleneck: not the stretching, the mixing.
Here is where AI earns its place. Source separation models — Demucs is the well-known one, MIT-licensed, developed by Alexandre Défossez at Meta — take a finished stereo mix and pull it back into drums, bass, vocals, and other. Once separated, you stretch each stem with the right settings: transient-preserving for drums, smooth spectral stretching for vocals and pads. Then recombine. The drums stay punchy, the vocal stays natural, and the result sounds like the band played slower rather than like the tape melted. The AI never touched the tempo — it just undid the mix so conventional DSP could do its job properly.
A note on the word "BPM." Consumer tools that promise to "change the BPM" are doing exactly this time-stretch. They aren't re-performing the song. If you want a genuine slower performance — different feel, different drummer choices — that's music generation, an entirely different tool and an entirely different rights conversation.
An example that makes it click
Think of a song as a train of dominoes falling.
Slowing the video makes everything lower-pitched and wrong — that's changing playback speed. Time-stretching is different: you go in and move each domino a little further apart, so the same falls happen, just spaced out. Pitch untouched.
But some of those dominoes are firecrackers (drums) and some are long organ notes (vocals). Spacing them with one rule wrecks somebody. So first you sort the firecrackers into one pile and the organ notes into another — that's the AI part — then you space each pile with its own rule, and set them off together. Same song, slower, and nothing sounds melted.
How to do it
- Decide what you actually want: slower playback at the same pitch (time-stretch — this is nearly always it), or lower pitch at the same speed (pitch-shift).
- For a quick pass, use a plain time-stretch: Audacity's Change Tempo, a DAW's warp mode, or a phone app's tempo slider. Try 5-15% first — small changes rarely produce artifacts.
- If it sounds watery, muddy, or the drums lose their snap, that's the mix fighting the algorithm. Move to stem separation.
- Run the track through a source separation model — Demucs (htdemucs is the default) or a hosted service built on it — to get drums, bass, vocals, and other as separate files.
- Stretch each stem with settings that suit it: transient/percussive mode for drums, spectral or complex mode for vocals and sustained instruments.
- Recombine the stems at the new tempo and check the drum hits against a click. Timing drift shows up in the drums first.
- If you're publishing the result, remember slowing a song is a derivative work — the recording and composition rights still apply.
Key facts
- Changing playback speed alters pitch and tempo together; time-stretching separates them and is standard DSP, not AI.
- Demucs separates a mixed track into drums, bass, vocals, and other, with an experimental 6-source model adding guitar and piano.
- Demucs is released under the MIT license and was developed by Alexandre Défossez at Meta.
- Demucs model variants include htdemucs (default), htdemucs_ft (fine-tuned), htdemucs_6s, hdemucs_mmi, and mdx/mdx_extra.
- Time-stretch artifacts arise because transients (drums) and sustained tones (vocals) require different stretching treatment, which is impossible on a single mixed track.
- Slowing down and re-publishing a copyrighted recording creates a derivative work and does not clear the underlying rights.
▶ The 60-second explainer (script)
You want a song slower, but not lower — no chipmunk, no monster voice. Here's what's really going on, and the surprise is that the slowing part isn't AI at all. Sound is pressure waves in time. Play a record at eighty percent speed and it gets longer and lower, because speed and pitch are welded together. The classic fix is called a phase vocoder: chop the audio into overlapping slices, and re-lay those slices further apart in time while leaving their frequencies alone. Longer, same pitch. That's been in Audacity and every DJ deck for years. No AI. So why does it sound like garbage? Because the trick assumes each slice is one steady thing, and music isn't. A drum hit is a sharp click. A vocal is a long tone. Stretch them with the same rule and you either smear the drums or turn the vocal underwater. The algorithm can't tell them apart. And that's where AI actually earns its keep — not stretching, unmixing. Models like Demucs, built at Meta and released under an MIT license, take a finished stereo mix and pull it back apart into drums, bass, vocals, and everything else. Now you stretch each one with the right setting: punchy for drums, smooth for vocals. Then recombine. It's dominoes: some are firecrackers, some are long organ notes. Sort them into piles first, space each pile its own way, then set them off together. The AI never touched the tempo. It just undid the mix so the old math could work.
What authoritative sources say
People also ask
Is slowing a song's BPM actually AI?
No. Time-stretching is conventional signal processing that predates modern AI by decades. AI's contribution is separating the song into stems first so the stretch sounds clean.
Why does my slowed track sound underwater?
You're stretching drums and vocals with one setting. Transients need different handling from sustained tones. Separate the stems, stretch each appropriately, then remix.
How much can I slow a song before it falls apart?
On a mixed track, roughly 5-15% is usually clean and beyond that artifacts creep in. With stem separation you can go considerably further before it sounds wrong.
Will slowing it down change the key?
Not with time-stretching — that's the entire point. Only a speed change moves the pitch. If you want a different key, that's a separate pitch-shift operation.
Can I post my slowed version online?
Slowing a copyrighted recording produces a derivative work. The recording and composition rights still apply, and platform rules and licensing don't disappear because the tempo changed.