音频和弦识别

Audio chord recognition, also called automatic chord estimation, takes a recording and produces a time-coded list of chord labels. A typical output reads like a lead sheet with timestamps: G:maj from 0.00 to 1.85 seconds, E:min until 3.70, C:maj until 5.55, and N, meaning no chord, during a drum break. A chord here is a set of pitch classes, the 12 note names regardless of octave, sounding together, described by a root such as G and a quality such as major, minor or dominant seventh. The task sits between full transcription, which writes every note, and key detection, which gives one label for a whole piece.

The classic pipeline has three steps. First the audio is turned into a chromagram, a sequence of 12-number vectors that measure how much energy falls on each pitch class in each short frame. A C major chord should light up C, E and G. Second, each frame is compared with chord templates, idealized chroma patterns for every chord in the vocabulary, or scored by a statistical model. Third, a temporal model smooths the frame-wise guesses, because chords usually last for beats or bars rather than flickering every 50 milliseconds. A hidden Markov model (HMM), which treats the chord as a hidden state that tends to persist and changes with learned probabilities, is the traditional choice, decoded with the Viterbi algorithm to find the most likely chord sequence.

Chordino, released by Matthias Mauch and Chris Cannam as a Vamp audio plugin, is a widely used example of this design. Its companion NNLS Chroma plugin converts the spectrum to a log-frequency scale, tunes and whitens it, and then performs approximate note transcription with non-negative least squares (NNLS), a fitting method that explains the spectrum as a sum of harmonic note shapes with no negative amounts. That cleans up overtones, which otherwise make a C note leak energy into G and E. Separate bass and treble chromagrams help identify the bass note. Chordino matches frames against chord profiles from a dictionary file and smooths them with HMM decoding. The method comes from a 2010 ISMIR paper by Mauch and Simon Dixon on identifying difficult chords.

The field's history is summarized in a 2019 ISMIR review by Johan Pauwels, Ken O'Hanlon, Emilia Gómez and Mark Sandler titled 20 Years of Automatic Chord Recognition from Audio. It traces the task to a 1999 paper by Fujishima and describes a shift from knowledge-based systems, built on music theory and hand-set templates, to data-driven methods, with neural networks now central. The authors argue that many of today's core issues were already raised in that first paper. Shared evaluation rests on hand-annotated corpora such as the Isophonics Beatles collection, whose chord syntax and transcription process are described in Chris Harte's 2010 PhD thesis; that syntax is now the standard way to write chord labels in research.

Deep learning replaced hand-built stages one at a time. In 2016 Filip Korzeniowski and Gerhard Widmer showed that a neural network fed a spectrum with surrounding context could learn chroma vectors that were less noisy than hand-crafted ones, improving recognition with even a simple linear classifier. In 2017 Brian McFee and Juan Pablo Bello trained a convolutional-recurrent network end to end on log-power constant-Q spectra, a frequency representation with bins spaced like musical notes. Their model used a 170-class vocabulary, 12 roots times 14 qualities plus N and X for unrecognized chords, and was also trained to predict the root, bass and active pitch classes as separate outputs. On 1,217 annotated tracks from the Isophonics, Billboard, RWC Pop and MARL collections it gained 4 to 5 percentage points in the hardest sevenths and tetrads categories over earlier convolutional and HMM systems. In 2019 Park and colleagues proposed a bi-directional Transformer for chord recognition, using self-attention to capture long-range context in a single training phase.

The honest measure of progress is sobering. McFee and Bello's best model identified the exact simplified chord, root and quality together, in 63.6% of frames. Their data explain why. By duration, major chords made up 52.5% of the corpus and minor chords 13.6%, while sixth chords were 1.5%, suspended chords 2.5% and minor-major sevenths only 0.03%. Models therefore lean toward major and minor, and they confuse a seventh chord with its plain triad. Out-of-vocabulary chords, mostly power chords and single notes, made up 2.2% of the data and were usually pushed toward major or minor. Simpler evaluations score far higher, because getting the root right or matching three of the notes is much easier than naming the full chord.

Other failure modes come from music itself. Annotators disagree on chord labels, especially for passing chords, implied harmony and dense jazz voicings, so the reference is partly subjective. Strong vocals, distorted guitars and drums blur the chromagram. The standard corpora, such as the Beatles, Billboard and RWC Pop collections, are mostly Western popular music, so other harmonic systems fit the labels poorly. Tuning that is not at standard pitch shifts energy between bins, which is why NNLS Chroma includes a tuning step. Inversions, where a note other than the root is in the bass, are often ignored; McFee and Bello's model predicted the bass but did not output inversions, while the later crema release added inversion tracking.

In practice, running a chord recognizer is cheap. Chordino runs on an ordinary computer inside Sonic Visualiser, and crema returns chord labels with confidence values from a single Python call. Evaluation is standardized through mir_eval, so results are comparable across papers if the vocabulary and comparison rule are stated. The useful habit is to report several scores, such as root, majmin and sevenths, and to listen to low-confidence sections rather than trusting a single accuracy number.