Automatic Melodic Transcription of Flamenco Singing

Background in ethnomusicology and music analysis. Melody in flamenco music is characterized by the presence ofmicrotonality, pitch glides (portamenti or smooth transitions), a short pitchrange or tessitura (usually limited to a range of a sixth and characterized bythe insistence on a note and its contiguous ones), the use of enharmonic scale(having microtonal intervallic differences between enharmonic notes), a Baroqueornamentation (with an expressive function) and the use of a Greek Dorian modein the most traditional songs (Fernández 2004).

Background in computing, mathematics and statistics. State of the art techniques in melodic analysis of audio allow us toobtain different representation levels of a music recording (Gómez et alt.2003A, Gómez et alt. 2003B). There are different representation levels formelody. Energy (computed for the whole spectrum or in different frequencybands) and fundamental frequency (related to its perceptual correlate pitch)curves are the main low-level melodic features. They represent theinstantaneous evolution of intonation and intensity. In a higher structurallevel, note duration and pitch provide a symbolic representation, which can bethe input to higher-level analyses. Finally, deviations of the analyzedrecording with respect to the obtained score are related to expressivity.

Aims. The goal of this work is to provide atool for the automatic melodic transcription of flamenco pieces in differentrepresentation levels, focusing on monophonic singing excerpts. The tool shouldhelp finding a suitable representation level for flamenco performances and beuseful for the comparative analysis of singers and styles.

Main contribution. The steps for melodicdescription can be summarized as follows. We first compute a set of low-levelfeatures from the audio signal, including energy and fundamental frequency.Next, since the tuning reference is unknown, we estimate it with a method basedon the equal-tempered chromatic scale. The audio signal is then segmented intonotes using a dynamic programming algorithm whose cost functions take intoaccount pitch error, energy variations and note durations. The obtained notesare labeled in terms of pitch and duration. Finally, vibrato depth and ratedescriptors are computed for each note. We have worked with a music collectionof 135 monophonic pieces corresponding to deblas, martinetes and tonás. Thesecategories of flamenco styles (called palos )are traditionally sung a cappella or in some cases with some percussion.

Results . The tool provides a way to visualize different aspects of theanalyzed performances, as the instantaneous evolution of energy and pitch, thesequence of notes and some expressive aspects (modulations and transitions).

Conclusions . The work presented here provides a set of techniques for theautomatic melodic description of flamenco singing. Quantitative evaluation is adifficult task, as it is necessary to have a set of manually annotatedperformances by expert flamencologists. This is an ongoing work, although someinformal tests have proven the usefulness of the approach. Actually, theproposed tool is being used in the context of a research project on comparativeanalysis of a cappella flamenco styles (Escobar et alt. 2007).

Implications . These techniques can support research on comparative analysis ofsingers and styles, and they provide an automatic and objective way totranscribe flamenco performances.