Approximate Note Transcription for the Improved Identification of Difficult Chords
Slides are available here.
Abstract. The automatic detection and transcription of musical chords from audio is an established music computing task. The choice of chord profiles and higher-level time-series modelling have received a lot of attention, resulting in methods with an overall performance of more than 70% in the MIREX Chord Detection task 2009. Research on the front end of chord transcription algorithms has often concentrated on finding good chord templates to fit the chroma features. In this paper we reverse this approach and seek to find chroma features that are more suitable for usage in a musically-motivated model. We do so by performing a prior approximate transcription using an existing technique to solve non-negative least squares problems (NNLS). The resulting NNLS chroma features are tested by using them as an input to an existing state-of-the-art high-level model for chord transcription. We achieve very good results of 80% accuracy using the song collection and metric of the 2009 MIREX Chord Detection tasks. This is a significant increase over the top result (74%) in MIREX 2009. The nature of some chords makes their identification particularly susceptible to confusion between fundamental frequency and partials. We show that the recognition of these difficult chords in particular is substantially improved by the prior approximate transcription using NNLS.