Improved Pitch and Voicing Determination Using the Reflected Root Chirp Group Delay Spectrum

Estimating the pitch accurately is an essential task in various audio processing applications, and methods based on deep learning are currently among the state-of-the-art. In this paper we present results using features derived using Reflected Roots Chirp Group Delay (RRCGD) spectrum and its autocorrelation. Compared with our previous work based on the ℓ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf>-norm residual, our best improvement in Raw Pitch Accuracy (RPA) was 3.31% and Voicing Recall Rate (VRR) was 5.02% in absolute terms and averaged over all SNRs. Incorporating the Transformer architecture, which captures context, helped in improving the accuracy. Unlike the ℓ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf>-norm residual approach, the proposed method can be used for estimating the pitch of music signals as well because it does not depend on the source-filter model.