Evaluation and applications of tonal profiles for automatic music tonality description

TitleEvaluation and applications of tonal profiles for automatic music tonality description
Publication TypeConference Paper
Year of Publication2006
AuthorsGómez, E., & Herrera P.
AbstractBackground
A large amount of research has been devoted to study how tonality is perceived, mainly in western music. Some of this research has lead to the development of algorithms for the automatic estimation of the key of a given piece by analyzing its score indicating that these tonal models can be used to estimate the tonality of unknown musical pieces. During the last few years, we have witnessed an increasing interest in analyzing audio recordings where the score is unknown (MIREX-05).

Aims
The aim of this work is discussing how a set of tonal profiles derived from human ratings can help to analyze pieces of music in audio format, where the score is unavailable.

Method
We evaluate the performance of different tonal models when applied to estimate the key of a piece, in comparison to the use of pitch class distribution profiles obtained empirically from statistical analysis and some flat profiles taken from music theory. We first obtain a representation of the pitch class distribution of an audio piece and then correlate these features with the studied tonal profiles in order to obtain an estimated global key measurement.

Two different models are compared the one by Krumhansl (Krumhansl-90) and the modifications by Temperley (99). We also include major and profiles derived from statistical analysis of folk MIDIs (Chai-05), from the Kostka and Payne (95) music theory textbook (Temperley-05), and some additional audio-derived statistics (Gomez-05). Finally, we include tonic triad and diatonic flat profiles derived from music theory.

Results and Conclusions
The use of flat diatonic, Krumhansl and Temperley profiles yields similar results. This performance is also similar to the use of statistical profiles obtained from audio features, but better than profiles empirically obtained from symbolic scores. In general, the performance degrades when analyzing musical genres different than classical music (e.g. pop, rock, etc). The best result for popular music (55% accuracy) is obtained using a tonic triad profile. This reveals that tonal models might be depending on the musical style and it supports the importance of the tonic triad in popular music (Temperley-01).

intranet