Analysis and Improvements of the Cepstrum Method for Fundamental Frequency Estimation in Music Signals
Johannes Gauer, Diana Kleingarn, Rainer Martin · 2021 29th European Signal Processing Conference (EUSIPCO) · 2021
Fundamental frequency estimation has many applications in speech and music processing. Among other methods, peak tracking in the cepstrum domain has been established as a robust and successful method for speech signals. Its elegance stems from the fact that a harmonic pattern in the log-spectral domain is mapped onto a single bin in the cepstral domain. However, the cepstrum method had not been thoroughly analyzed in the context of music signals. In this work we introduce a novel method for detecting and compensating octave errors to enhance the fundamental frequency search in the cepstrum and combine it with a fine search based on a least squares approximation in the time domain. The performance of the method is evaluated with monophonic music signals across a wide range of fundamental frequencies. In addition, the achievable frequency resolution and estimation error and their dependence on general signal parameters are analyzed.