Abstract: Accurate pitch and voice detection is fundamental in speech and audio processing, and plays a significant role in synthesis, speech recognition, and audio analysis. In noisy conditions, effective pitch and voicing estimation contribute to better speech enhancement and separation. Traditional approaches rely on signal processing principles whereas neural network-based methods are based on deep learning. In our research, instead of using the raw waveform, we have investigated alternative representations for enhancing the robustness of pitch estimation.

In particular, our features are based on the Reflected Roots Chirp Group Delay (RRCGD) spectrum and its autocorrelation. The proposed approach was evaluated on the PTDB-TUG dataset. For the 0dB SNR condition, our method improved Raw Pitch Accuracy (RPA) by 3.31% and Voicing Recall Rate (VRR) by 5.02% in absolute terms, averaged across all SNR levels, compared to our previous L-1-norm residual-based approach. The inclusion of the Transformer architecture, which effectively captures temporal context, contributed to these improvements. Additionally, unlike L-1 norm residual methods that rely on the source-filter model, our approach is also applicable to pitch estimation in music signals, broadening its usability beyond speech processing.

Event Details
Title: Improved Pitch and Voicing Determination Using the RRCGDS Algorithm.
Date: July 07, 2026 at 03:00 PM
Venue: Google Meet (https://meet.google.com/fbu-dwtj-mye)
Speaker: Mr. NISHANT SINGH (EE22S084)
Guide: Dr. Ramalingam C S
Type: MS seminar

Updated: