Predominant Fundamental Frequency Estimation vs Singing Voice Separation for the Automatic Transcription of Accompanied Flamenco Singing

This work evaluates two strategies for predominant fundamental frequency (f0) estimation in the context of melodic transcription from flamenco singing with guitar accompaniment. The first strategy extracts the f0 from salient pitch contours computed from the mixed spectrum; the second separates the voice from the guitar and then performs monophonic f0 estimation. We integrate both approaches with an automatic transcription system, which first estimates the tuning frequency and then implements an iterative strategy for note segmentation and labeling. We evaluate them on a flamenco music collection, including a wide range of singers and recording conditions. Both strategies achieve satisfying results. The separation-based approach yields a good overall accuracy (76.81%), although instrumental segments have to be manually located. The predominant f0 estimator yields slightly higher accuracy (79.72%) but does not require any manual annotation. Furthermore, its accuracy increases (84.68%) if we adapt some algorithm parameters to each analyzed excerpt. Most transcription errors are due to incorrect f0 estimations (typically octave and voicing errors in strong presence of guitar) and incorrect note segmentation in highly ornamented sections. Our study confirms the difficulty of transcribing flamenco singing and the need for repertoire-specific and assisted algorithms for improving state-of-the-art methods.