Performance Improvement and Analysis of a Method for Piano Authentication From Audio Signal by Combining Feature Vectors of LPC Spectral Envelope, MFCC, and pLPC Pole Distribution

This paper presents a method for performance improvement by combining feature vectors in piano authentication from the audio signal. So far, we have shown that the combination of the linear predictive coding spectral envelope (LPCSE), the Mel-frequency cepstral coefficients (MFCC) and the piecewise linear predictive coding pole distribution (pLPCPD) improves the performance in speaker verification or authentication, where pLPCPD is the feature vector that we have introduced and developed for speaker verification. We also have analyzed the performance improvement from the point of view of the aperiodicity extracted by pLPCPD and the periodicity extracted by LPCSE and MFCC. This paper applies the method to the verification or authentication of three piano makers: Yamaha, Bösendorfer, and Steinway. Different from speaker verification, a piano has 88 keys to produce a wide range of pitch sounds and has several play styles, such as normal, staccato, tremolo (repeated notes), and pedal. Through the experiment of piano maker verification using all datasets involving different pitches and play styles, we show that pLPCPD+LPCSE (the combination of pLPCPD and LPCSE) has achieved the best performance. Through the experiments using restricted datasets, we show that pLPCPD+MFCC+LPCSE and pLPCPD+MFCC have achieved the best performance in the pitch and the play-style dependent verification, respectively, while pLPCPD+MFCC and pLPCPD have achieved the best performance in the pitch and the play-style independent verification, respectively. As a result, pLPCPD has the largest amount of information dependent on the piano makers and independent from the pitch and the play-style, while the three features have supplementary information each other.