SongF0: A Spectrum-Based Fundamental Frequency Estimation for Monophonic Songs

Abstract

In this paper, we propose a fundamental frequency ( $$f_0$$ ) estimation method for monophonic songs. The proposed method (SongF0) performs simultaneous voiced/unvoiced detection (VUD) and $$f_0$$ estimation from the frequency spectrum. Even though the spectrogram exhibits ample information of the singer $$f_0$$ in the form of harmonic partials, the existing spectrum-based $$f_0$$ detection methods fail to accurately extract the $$f_0$$ due to the complex singing styles. In order to cover a wider range of fundamental frequencies, singers adjust their vocal tract shape by lowering the larynx and widening the pharynx results in tuning lower harmonic partial to formant frequencies (Sundberg in STL-QPSR 1:1–6, 1968; Speech Transm Lab Q Prog Status Rep 4:21–39, 1970). Hence, most of the available popular $$f_0$$ detection methods which are inspired by speech production mechanism are susceptible to the energized higher-order harmonic spectral partials near the formants. In this work, we profoundly explore the quasi-harmonic nature of the spectral peaks to minimize the effect of formant frequencies on the detected $$f_0$$ . Initially, we train an ensemble classifier with the novel spectral features to predict the candidate $$f_0$$ harmonic partials from the spectrum. We exploit the property of constant spectral distance between the harmonic partials to reliably extract the singing $$f_0$$ from the predicted harmonic partials. We propose novel post-processing methods to significantly improve the $$f_0$$ detection accuracy in the weakly voiced and inharmonic transition regions. The proposed SongF0 which is independent of the vocal $$f_0$$ range is compared with the state-of-the-art $$f_0$$ extraction methods proposed for both speech and singing voice. The evaluation results on the openly available singing $$f_0$$ gold standard datasets revealed that the proposed method is significantly better than the several state-of-the-art $$f_0$$ detection methods.