Beat Sync Video: Cut to the Beat of the Music Automatically

Try Beat-Aligned Cutting Free

Upload footage, add a track, type one sentence. 50 credits, no card.

How to Cut a Video to the Beat of the Music

1. 1
   Upload your footage and add the music

   Drag in any MP4, MOV, MKV or WebM up to 14GB or 3 hours. Then add the track: "add the driving electronic track from the library", upload your own file, or paste a link. Music has to be in the edit before you align, because the beat you are cutting to is the song the viewer hears — not the footage's own room tone.
2. 2
   Make the cuts you want

   Beat alignment moves existing cuts, so give it cuts to move: "cut the boring parts", "keep the four best moments", or "build a montage from the shots where I'm outside". A single uncut span has no internal junctions and nothing to align.
3. 3
   Type "align the cuts to the beat"

   The agent measures the track, reports the tempo and its confidence, and snaps every junction that has a beat within tolerance. It tells you exactly how many moved, how many were already on the beat, and how many it skipped and why.
4. 4
   Review, adjust, export

   Watch the preview. Ask for a wider tolerance, a specific BPM, or a nudge on a single junction — "pull the cut at 0:14 back three frames". When it feels right, export a full-quality H.264 MP4 rendered from your original file.

Analysis of the music runs once per track and is cached. Every alignment pass after that is a chat message.

What Alignment Actually Does to a Cut

Three things can happen to a junction when the grid is applied. Only one of them is a move, and the other two are refusals worth understanding — they are why beat-aligned edits made this way still hold together.

How Beat Alignment Works Underneath

There are two hard problems here, and most tools only solve the first one. Where are the beats, and which cuts are allowed to move onto them?

Finding the beats

The audio is decoded and turned into a spectral-flux onset envelope: for each short frame, the sum of how much energy has increased in each frequency band since the previous frame. That measures attacks — a kick, a snare, a plucked string — rather than loudness, which is why a track that swells does not read as a hundred beats. The envelope is trend-removed and its statistical outliers are clipped, so one freak transient cannot dominate the analysis that follows.

Tempo comes from autocorrelating that envelope: the lag whose repeats line up best with the onsets is the beat period, and 60 divided by it is the BPM. That correlation strength becomes a confidence score, and it is the number that decides whether alignment happens at all — below 0.5 the tool refuses, on the principle that a synced-looking edit built on an imaginary pulse is worse than no sync.

The grid itself is then phase-locked: 64 candidate phases are tried and the one whose beats collect the most onset energy wins. Each beat is finally refined toward the strongest local peak within about 15% of the beat period, but only if that peak clearly beats the grid position. Both halves matter. A rigid mathematical grid drifts away from human-played music within thirty seconds; a pure peak-picker follows the loudest hits and loses the meter entirely. Refining a grid with a bias back toward the grid does neither.

The music is the beat, not the footage

When there is music in the edit, that track is what gets measured. This sounds obvious and it is the thing that most often goes wrong: a montage over a song still has its own audio — footsteps, wind, a bottle on a table — and those transients form a rhythm that has nothing to do with what the viewer hears. Aligning to them is aligning to the wrong sound entirely. The footage's own audio is available as a deliberate choice, which is the right one for a live performance or a drum take recorded in camera, and nowhere else.

The 23 tracks in the built-in royalty-free library carry a tempo measured offline across the whole file by a stronger estimator than a single agent turn can afford. When the in-turn estimate of a library track scores weaker than that catalogue figure, the catalogue wins — one real montage was once told its own soundtrack had no beat because a streamed excerpt scored 0.44 where the full-file measurement scored 0.62.

Two clocks, and why only one of them counts

A beat happens at a moment in the finished video. A cut is stored as a boundary in the original footage. Those are different clocks, and the conversion between them is the part that makes naive implementations fail. What the viewer hears at a junction is decided by how much footage precedes it, so the thing that has to move is the end of the span before the junction — and moving it shifts every junction after it.

So junctions are processed left to right and the timeline is recomputed after each move. A single pass over source timestamps cannot do this correctly, which is why beat sync built on "nudge each clip a bit" drifts progressively worse toward the end of a sequence. Any speed change in the way is handled in the same conversion: a move of d seconds in the finished video becomes d multiplied by the playback factor in source seconds.

The three refusals

Tolerance. Default 0.35 seconds, adjustable from 0.05 to 1.0. A junction with no beat inside that window is left where it is, because dragging a cut a full second to reach a beat destroys the reason you put it there. Worth doing the arithmetic: at 120 BPM the beats are 0.5s apart, so nothing is ever more than 0.25s from one and the default tolerance never rejects anything. Tolerance only starts refusing below roughly 86 BPM, where half a beat period exceeds 0.35s.

Words. Every candidate is checked against the word-level transcript built during analysis. If the beat falls inside a spoken word, the move is skipped. Eighty milliseconds off the beat is a rhythm choice nobody will name; a cut through the middle of "everything" is a defect everybody hears.

Collisions. A move is refused if it would invert a span or close the gap after it. A removed filler word is often a gap roughly twice the default tolerance wide — without a margin, both of its edges could snap to the same beat and silently restore footage you deliberately cut. The program's very first frame and very last frame never move at all.

Every pass reports the counts: how many moved, how many were already on the beat, how many were out of range, how many hit the word veto. If nothing moved, the result says so in those words, and Valmera's honesty layer verifies the reply against what the system actually recorded — so the agent cannot tell you the cuts are beat-aligned when they are not.

BPM, Beat Grid, Downbeat, Phrase

Four words that get used interchangeably and are not interchangeable. BPM is beats per minute; divide 60 by it for the beat period, so 120 BPM is a beat every 0.5 seconds and 90 BPM is a beat every 0.667. The beat grid is the list of moments those beats land on — that is what alignment snaps to. The downbeat is beat one of each bar; in 4/4 that is every fourth beat, and it is where the music itself says a new unit starts. A phrase is usually four or eight bars — sixteen or thirty-two beats — and it is the level at which a listener feels sections change.

Be clear about what the analysis produces: a beat grid, and a single tempo. It does not detect downbeats, count bars, infer a time signature, or track a tempo that changes mid-song. So the musical hierarchy is not something the snapping knows — it is something you express by deciding where the cuts go before you snap them.

Which brings up the mistake that makes beat-synced edits look amateur: cutting on every beat. At 120 BPM that is 120 cuts a minute, one every half second. A viewer needs a meaningful fraction of a second just to find the subject of a new composition, so at that rate the shots stop being read as content and start being read as texture. It can be exactly right for eight seconds as a build; it is exhausting for ninety.

The version that reads as professional is hierarchical, matching the music. Cut the body on downbeats — every fourth beat in 4/4 — so each shot gets a bar to breathe. Put your structural cuts on phrase boundaries, every four or eight bars, where the track changes anyway; a cut there is invisible in the best sense, because the music already told the viewer something was about to change. Then drop to every beat, or every half beat, for one short burst at the peak, and go back. Valmera's audio analysis reports the loudest moment of the track, the quietest, and the largest energy rise with the point where that rise resolves, which is how you find those boundaries without counting bars by ear.

On the Beat, or a Few Frames Early?

The convention among editors is that a cut often lands better two or three frames before the beat than exactly on it. There is a reason it works: the eye needs a moment to find the new subject, so a shot that arrives fractionally early has already been parsed by the time the transient hits, and the hit feels like it belongs to the new image. Cut precisely on the beat and the viewer is still locating the frame when the sound arrives, so the two events register as sequential rather than simultaneous.

It is not a universal rule. When the incoming shot contains the hit — a stick striking a drum, a hand landing on a table, a foot hitting the ground — you want the cut on or barely before the beat, because arriving too early shows the anticipation and gives the impact away. And when the outgoing shot has strong motion, a couple of frames late can read better than early.

Practically, in Valmera: alignment puts cuts on the grid, and there is no lead or lag parameter. Run the alignment first, watch it, and then nudge the specific junctions that want it — "pull the cut at 0:14 back three frames". That order matters. Three frames is about 100 milliseconds, comfortably inside the default tolerance, so a second alignment pass would snap your nudged cut straight back onto the beat.

When It Goes Wrong

How People Cut to the Beat Without Valmera

Beat-cutting is old craft and there are good tools for it. Worth knowing them, because for some workflows they are the better answer.

DaVinci Resolve detects beats natively: right-click a music clip on the timeline and enable Show Music Beats, and Resolve analyzes the track and draws beat markers along it. With snapping on, clips you trim or move land on those markers. Check your edition before planning around it — Show Music Beats arrived in Resolve 20 and sits on the paid Studio side rather than the free build. Studio is a one-off licence rather than a subscription, and if you already own it this is the strongest desktop route.

Premiere Pro has no built-in beat detection. The standard workaround is to play the track and tap M in time with it to lay markers by hand, which is fast, surprisingly accurate, and completely serviceable. The paid answer is BeatEdit, a third-party plugin from mamoworld that detects beats, writes the markers, and can generate an edit in sync with them. Premiere's Remix feature is a different job: it retimes the music to fit your video's length rather than fitting your cuts to the music.

CapCut auto-detects beats on the audio track and drops markers you can snap clips to, on both mobile and desktop, and it is the quickest of the free consumer routes. The trade is the usual one: the templates that make it fast also make the result look like everything else made that way.

ffmpeg cannot do this by itself — it has no beat detection at all. The command-line route is to get beat times from a dedicated analyzer such as aubio's aubiotrack or librosa's beat tracker, generate a cut list from them in a script, and feed that to ffmpeg. It is genuinely good if you want beat-cutting inside a pipeline, and it is exactly the machinery Valmera runs for you.

What every one of these gives you is beat markers. Deciding which beats deserve a cut, moving your existing junctions onto them, and doing it without slicing a word in half or quietly restoring footage you removed is still manual work in all of them. That last part is the whole difference: Valmera moves the cuts, and refuses the moves that would damage the edit.

Honest Limits

• Alignment moves cuts, it does not create them. For a montage from scratch, ask for the spans to be built on beat times first — then snapping is a refinement, not the edit.
• The analysis yields a beat grid and one tempo. No downbeat detection, no bar counting, no time signature, no tempo-change tracking — a song that shifts tempo halfway gets a single grid.
• Below 0.5 confidence it refuses. That is deliberate and it will not be argued out of it — but it will use a tempo you supply, because your ears outrank the estimator.
• Cuts land on the beat. There is no lead or lag offset; the few-frames-early feel is a follow-up nudge.
• Tolerance caps at 1.0 second, and the program's first and last frames never move.
• The longest music item in the edit drives the grid. Layering several tracks does not produce several grids.
• Transitions are one style applied to every cut — flash is the one built for beat-synced cutting — and there is no true crossfade or per-cut choice.
• There is no AI music generation. Bring a track, paste a link, or use the 23-track CC0 library.

What Pairs With Beat-Aligned Cutting

A beat-cut sequence is usually the middle of a job rather than the whole of it. The same chat can add and duck the music, make the cuts you want to align, punch in on the moments that carry weight, drop one beat into slow motion, and reframe the result to 9:16 before export. If the piece is a short cut down from something long, the clip maker is the better starting point.

The mechanics of music, ducking and loudness are documented in music & audio; speed, zooms and transitions in speed & motion. Every tool has a page in the tools hub, and 50 prompts that work is the fastest way to see how requests like these are phrased.

Put the Cuts on the Beat

Add the music, make the cuts, type one sentence. The agent reports exactly what moved.

Frequently Asked Questions

Beat Sync Your Next Edit

The agentic AI video editor — describe the edit, review the preview, download a full-quality export. 50 free credits, no card.