Song Practice Studio, Explained
Drop a song into the studio and it returns the tempo, the key and a chord on every beat, then lets you loop, slow down and transpose it to practise along. This article explains how the machine listens, how to read what it gives back, and where its guesses go wrong.
What the studio actually does
Give it a recording and it answers three questions in order. How fast is it? Where are the beats and bars? What chord is sounding on each beat? From those answers it builds a chord sheet that scrolls with the audio, draws a shape for guitar, ukulele or piano under each chord, and offers a capo position that turns awkward keys into open shapes.
Here is a concrete hook. Feed it a four-chord pop song in A major at 120 beats per minute. The studio reports 120 BPM, labels the key A major, and lays out the progression A - E - F#m - D, one chord per bar. The F#m is a barre chord, so it suggests a capo on fret 2, after which the same four chords are played as G - D - Em - C, open shapes most beginners already know. The audio does not change pitch; only the names and shapes change to match where your fingers go.
Everything runs in your browser. Nothing is uploaded, so the number of songs you study is limited only by your patience.
When to use it, and when not
Use it to learn a song by ear faster, to find a capo position, to build a backing track you can slow down, or to loop a tricky four bars until your hands catch up. It is strongest on pop, rock and folk, where chords change on the bar line and the harmony is plain triads and sevenths.
Do not treat the output as a transcription. Expect most beats in a straightforward song to be correct and a handful to need your ear. The decoder knows 96 chords: major, minor, dominant seventh, minor seventh, major seventh, sus2, sus4 and diminished on all twelve roots. A song built from jazz extensions (a Cmaj9 or a G13) will be rounded to the nearest shape it knows. Heavily distorted guitar smears its overtones across the spectrum and confuses the pitch estimate, so metal rhythm parts are the hardest case.
If the chord sheet counts twice too fast, or every chord looks shifted by half a bar, the beat grid is off, not the chords. Use the Fix the beat grid buttons to halve the tempo or nudge the bar start, then re-read the sheet. Fixing the grid usually fixes the chords in one step.
How the listening works
The first stage is an onset envelope: a curve that rises wherever the audio gets suddenly louder or brighter, which is where notes and drum hits begin. Peaks in that envelope repeat at the tempo. If strong peaks land every 0.5 seconds, that is two per second, or 120 beats per minute.
Here T_{\text{beat}} is the time in seconds between one beat and the next. At T_{\text{beat}} = 0.5 you get 60 / 0.5 = 120. At T_{\text{beat}} = 0.6 you get 100. A beat tracker then places each beat on a peak, keeping the spacing close to the tempo while following a drummer who drifts.
The second stage reads pitch. On each beat the studio measures how much energy sits at each of the twelve pitch classes (C, C#, D, and so on), folding every octave together into a 12-number chroma vector. Before measuring, it suppresses overtones: a plucked low E rings not only at E but at B and G# above it, and those harmonics would otherwise vote for the wrong chord.
Each beat's chroma vector is compared against 96 templates. A C major template puts most of its weight on C, E and G, plus a little on G, B and D, the notes a fifth above each chord tone, where leftover overtones land because no overtone filter is perfect. The template with the best match, nudged up when the bass note agrees with the root, is the raw chord guess for that beat.
Why the bar line matters
Raw per-beat guesses are noisy. A sustained C major chord might match C on beats 1 and 3 but drift to Am on beat 2 because the melody briefly lands on A. Choosing each beat independently would produce a jittery, wrong sheet.
So the studio does not choose beats independently. It runs a decoder that scores a whole sequence at once, trading off two things: how well each chord matches its beat, and a penalty for changing chord. The penalty is smaller on a bar line than inside a bar, because real songs change chords on the downbeat far more often than mid-bar.
The first sum rewards chords that fit the audio on each beat t. The second sum counts chord changes, and \lambda is how much each change costs. The bracket is 1 when the chord changes and 0 when it holds. A larger \lambda gives longer, steadier chords; a smaller one lets the sheet follow quick changes. Because the cost is lower on downbeats, the decoder prefers to place changes there, which is why your sheet lines up with the bars. A small bonus for chords that belong to the detected key works the same way.
Reproducing the demo song
The demo is a song synthesized in your browser: a strummed guitar, a bass, drums and a melody in G major. Follow its numbers through.
- The onset peaks repeat every
0.652seconds. From 60 / 0.652 the tempo is92 BPM, and with 4 beats per bar each bar lasts about2.6seconds. - Over the first bar the chroma vector is heaviest on G, then B and D (chart below). That matches the
Gmajor template best, so bar 1 isG. - Bars 2 to 8 match
D,Em,C,G,D,CandD7, and the eight-bar loop repeats. Every chord belongs to G major, and the key estimate agrees:G major. - These are all open guitar shapes, so there is no capo tip. Transpose the song up one semitone for a singer and the chords become
Ab, Eb, Fm, Db, all barre chords; the capo tip then suggests fret 1, which turns them back into the G, D, Em and C shapes.
The chord sheet reads G | D | Em | C and G | D | C | D7, one chord per bar, scrolling about 2.6 seconds per bar, with the shapes underneath.
Reading and interpreting the results
Three outputs deserve a careful eye.
- Key
- The tonal center the decoder assumed. It gives a small bonus to the chords that belong to it and decides how sharps and flats are spelled. The usual slip is the relative key, E minor instead of G major: the two share the same chords, so the sheet barely changes, and when the evidence is close the studio shows the alternative next to the key.
- Chord per beat
- A best match, not a certainty. Where the sheet holds one chord for a whole bar, trust it. Where it flips chord every beat, be suspicious: that often signals a passing chord, a bass walk, or a weak pitch estimate.
- Capo tip
- The fret that yields the most open shapes. The chord names then describe the shapes you finger, not the sounding pitches. A capo-4 sheet showing
Csounds asE.
Practising with the result
Slowing down uses the browser's own pitch-preserving playback, so a drop to 70% speed is instant and the key is unchanged. Transposing and the vocal controls are different: they render a new copy of the audio in the background, which takes a moment on a long track.
The vocal controls look for what is identical in the left and right channels. A singer mixed dead center reaches both speakers with the same strength and timing, so the studio splits the sound into narrow frequency bands, measures in each one how alike the two channels are, and turns down only the parts that match. It leaves everything below roughly 90 to 180 Hz alone, so the bass and the body of the kick drum survive, but anything else panned center in the vocal range, typically the snare, softens with the voice. On a mono recording both channels are identical, so the control is switched off.
To build speed, loop four bars, start at 75% and turn on the speed-up: each pass adds 5%, so the sixth pass plays at full tempo without a jump you would notice. Add a click on the beat while you learn the changes, then switch it off once your timing holds.
Common mistakes
The most frequent error is reading a grid problem as a chord problem. If every chord seems shifted or doubled, fix the beat grid first. The second is trusting rapid chord flicker: a real song rarely changes chord every single beat, so a stretch of C Am C Am on four consecutive beats is more likely one held chord that the pitch estimate could not settle. The third is forgetting the capo convention: when a capo is suggested, the names are shapes, and the exported MIDI and the sounding audio stay at concert pitch.
One more: distorted or bass-heavy mixes blur the chroma. If a song fights the detector, switch on Simple chords to fold the sevenths and suspensions into plain major and minor, then correct what remains by ear.
Exports
You can take the work out in four forms. A plain-text chord sheet reads in any editor. A ChordPro grid drops into songbook apps. A MIDI file carries the chords with a tempo map that follows the recording beat by beat, so it drifts exactly as the original does. The transposed or vocal-reduced audio exports as WAV. All of it is generated locally.
Related tools
If you want to see the audio rather than its chords, the Waveform & Spectrogram Viewer shows the same signal as a waveform and a spectrogram, which is the clearest way to spot distortion smearing the harmonics. If you only need the tempo, or you suspect the beat grid counted double, the BPM Tapper & Detector lets you tap a tempo by hand and compare it against the detected value.
Frequently asked questions
Why are some chords wrong?
The studio matches each beat against 96 shapes and smooths the result, which is a strong first draft, not a transcription. Passing chords, jazz extensions and distorted guitars are the usual failures. Fix the beat grid first, then correct remaining beats by ear.
Does slowing the song change its pitch?
No. Slowing uses pitch-preserving playback, so 70% speed sounds the same key, just slower. Transposing is the separate control that changes pitch.
Why did removing the vocals also soften the drums?
Vocal removal turns down whatever sits in the center of a stereo mix in the vocal range. The snare is usually panned center too, so it thins out along with the voice; the bass and the low thump of the kick are left alone. Telling a centered voice from a centered snare needs a source-separation model, which this method does not use.
What does the capo tip actually change?
Only your fingering and the chord names printed on the sheet. It finds the fret that produces the most open, barre-free shapes. The sounding pitch of the recording and the exported MIDI stay the same.
Are my files uploaded anywhere?
No. All analysis, playback and export happen in your browser. Nothing leaves your device.