Fuzzy Analysis in Pitch-Class Determination for Polyphonic Audio Key Finding

This paper presents a fuzzy analysis technique for pitch class determination that improves the accuracy of key finding from audio information. Errors in audio key finding, typically incorrect assignments of closely related keys, commonly result from imprecise pitch class determination and biases introduced by the quality of the sound. Our technique is motivated by hypotheses on the sources of audio key finding errors, and uses fuzzy analysis to reduce the errors caused by noisy detection of lower pitches, and to refine the biased raw frequency data, in order to extract more correct pitch classes. We compare the proposed system to two others, an earlier one employing only peak detection from FFT results, and another providing direct key finding from MIDI. All three used the same key finding algorithm (Chew’s Spiral Array CEG algorithm) and the same 410 classical music pieces (ranging from Baroque to Contemporary). Considering only the first 15 seconds of music in each piece, the proposed fuzzy analysis technique outperforms the peak detection method by 12.18% on average, matches the performance of direct key finding from MIDI 41.73% of the time, and achieves an overall maximum correct rate of 75.25% (compared to 80.34% for MIDI key finding).