A Cognitive Method for Musicology Based Melody Transcription

Abstract

This paper describes a method for transcribing the main structure of polyphonic music audio automatically by analyzing musical tonality related musicological information. Music transcription is a difficult topic in Music Information Retrieval (MIR) which contains many tasks to recognize all the elements of music. Among all these musical factors, tonality is a basic but elusive one, and plays as the core in organizing music structure. So by making an investigation into cognition and musicology, we propose an approach to transcribe major component of polyphonic music from the point of view of tonality. Unlike previous studies, our method focuses on music cognition theory instead of signal processing and pattern classification. Basing on our former work on key finding and chord recognition, we introduce a new cognitive feature called auditory saliency (AS), which contains both statistically modeled information about human’s acoustic attention and psychology measured data on human’s musical perception, to recognize the main part of the music—melody stream. Constant Q transform which is a more “musical” STFT (Short-Time Fourier Transform), and a temporal ANN (Artificial Neuro Network) are also used in our framework. There are also a few musicology and signal processing based techniques designed to improve our method such as melody structural vertical regulating and onset detection. Our proposed method has been tested objectively (F-measure of pitch detection is 0.76 with a perfect musicological AS information) and subjectively (above 90% of transcribed segments are accepted by professional reviewers), and shows some inspiring results.