Neither of these choices would be objectively wrong, but each expresses a subjective selection of the harmonic content of the audio signal.įurthermore, reharmonization (altering an original harmony) is a common phenomenon in harmony transcriptions of popular music, which can happen implicitly because of perceptual differences between annotators, or explicitly to make a transcription more useful in a particular context. For example, if in a recording the simultaneously sounding notes C, E, and G are combined with a melody touching a B, it is up to the annotator whether to include B in the chord label ( C:maj7) or not ( C:maj). The harmonic content of an audio recording is often ambiguous and can result in annotators disagreeing about which chord label best describes a musical segment. The problem of “ground truth” in harmony transcriptionsĪnnotators transcribing chords from a recording by ear can disagree because of personal preference and bias toward a particular instrument, and because harmony can be ambiguous perceptually as well as theoretically by definition . argue, is that the perception of chords in recorded music can be highly subjective, which is problematic for deriving a single reference “ground truth” chord label annotation. Although current state-of-the-art ACE performance power allows them to be used in commercial products (e.g., Chordify, Footnote 1 Riffstation Footnote 2), their performance nevertheless seems to be tapering off in recent years Footnote 3 . Both feature extraction and pattern matching in modern ACE systems are commonly performed using machine learning technique in current state-of-the-art ACE systems usually some flavor of deep learning.
ACE systems consist of some variation of audio feature extraction followed by a pattern matching step in which the audio features are associated with chord labels. Our results show that annotator subjectivity should inform future research on automatic chord estimation to improve the state of the art.Įxtracting time-aligned sequences of chords from a given audio music signal, commonly referred to as automatic chord estimation (ACE), is a well-researched topic in music information retrieval (MIR). Furthermore, we show that chord personalization using multiple reference annotations outperforms using just a single reference annotation. From a single trained model and the annotators’ chord-label vocabulary, we can accurately personalize chord labels for individual annotators. We integrate these representations from multiple annotators and deep learn them from audio.
We introduce the first approach to automatic chord label personalization by modeling annotator subjectivity through harmonic interval-based chord representations.
Recent studies suggest that subjectivity is intrinsic to harmonic reference annotations that should be embraced in automatic chord estimation rather than resolved. Nevertheless, theoretical insights on harmonic ambiguity from harmony theory, experimental studies on annotator subjectivity in harmony annotations, and the availability of vast amounts of heterogeneous (subjective) harmony annotations in crowd-sourced repositories make the notion of a single-harmonic “ground truth” reference annotation a tenuous one. Current automatic chord estimation systems are trained and tested using datasets that contain single reference annotations, i.e., for each corresponding musical segment (e.g., audio frame or section), the reference annotation contains a single chord label.
0 kommentar(er)
