SongUp AI 免費和絃辨識工具在瀏覽器中能聽出甚麼、不能聽出甚麼

I build SongUp AI, and one of the free tools on the site is a chord finder: you drop in an MP3, WAV, M4A or FLAC file and it returns the chord progression with a timestamp for every change, plus the key and the BPM. The part I care about most is where it runs. The analysis happens in your browser, so the song never leaves your device. That choice shapes what the tool is good at, and it is worth explaining honestly.

What the tool actually does

At a high level there are two jobs. The first is finding when things change. Onset analysis looks for the moments where energy jumps, which is where notes and chords tend to start. Those onsets also give the tempo estimate.

The second job is finding what is playing. A chromagram folds every frequency in the audio into twelve pitch classes, C through B, regardless of octave. A C major chord lights up C, E and G. The tool compares each slice of the chromagram against chord profiles, 168 chord types in all: major, minor, sevenths, diminished, augmented and suspended chords. The best match per segment becomes a chord on the timeline. Summing the chroma across the whole song gives the key.

Why in the browser

Three reasons. Privacy: people upload unreleased demos and other people's songs they are learning, and neither needs to sit on a server. Cost: running the analysis on the visitor's device means the tool can stay free with a free account and no daily limit. Speed: there is no upload step for a 40 MB WAV, so the wait is only the analysis.

The cost of that choice is that the model has to be light enough to run on a phone, so it will not match a trained musician on dense material.

Where it works well, and where it does not

It works best on songs where the harmony is clear: a guitar or piano with a voice, pop with a steady band, a singer-songwriter demo. It struggles where any chord detector struggles:

• Heavy distortion and dense mixes. Distorted guitars add many overtones, which smear the chroma.
• Bass-less or bass-heavy mixes. Inversions (C/E) are hard to separate from a different chord.
• Fast passing chords. A chord that lasts half a beat may merge with its neighbour.
• Extended jazz voicings. A 13th chord can read as a simpler chord with the same core notes.

The practical fix for the first two is the one the page suggests: if you can, split the song into stems first and run the chord finder on the instrument stem without the vocal. Less going on means a cleaner chromagram.

How I use the result

I treat the output as a first draft. I copy the progression as text, play along with the track, and fix the two or three chords that sound wrong. That is still much faster than working out a whole song by ear, and the timestamps make it easy to jump to the bar that needs checking.

If you want to try it on a song you are learning: AI Chord Finder: chords, key and BPM of any song.

I would like to hear which songs trip it up, especially in genres outside pop, so I know where to look next.

Drafted with help from an AI writing assistant and edited by me.