Abstract
Fundamental frequency ($F_0$) estimation is a core component of speech technology, yet its robustness in reverberant environments remains a challenge. This paper investigates how physical room acoustics, specifically reverberation time ($T_{60}$) and speaker-to-microphone distance (SMD), degrade pitch tracking. By evaluating several algorithms across diverse datasets, we demonstrate that male voices are significantly more susceptible to reverberation-induced errors than female voices. Our results indicate that $T_{60}$ is an insufficient metric for predicting performance, as the Direct-to-Reverberation Ratio (DRR), dictated by SMD, plays a dominant role in the "far field."
Introduction
Identifying the "pitch" or fundamental frequency ($F_0$) of a human voice is essential for applications ranging from hearing aids to automatic speech recognition (ASR). While contemporary algorithms achieve high accuracy in "dry" or noise-free environments, their performance often collapses in real-world settings where sound reflects off surfaces. Traditionally, researchers have characterized these environments using reverberation time ($T_{60}$)—the duration required for sound to decay by 60 dB. However, this study posits that $T_{60}$ fails to capture the full complexity of the problem, particularly the critical interaction between the speaker's distance from the microphone and their vocal characteristics.
Methodology
The study compared a wide array of $F_0$ estimation algorithms, including traditional methods like Autocorrelation (ACF), Cepstrum, and Subharmonic Summation (SHS), alongside modern standards such as YIN and Normalized Cross-Correlation (NCCF). Specialized approaches like TEMPO and Complex Cepstrum Analysis (CCA) were also included.
To ensure robust findings, the researchers utilized two primary datasets:
- Japanese Database: 840 utterances from 28 speakers.
- UK English Database (TC-STAR): High-precision recordings with ground-truth pitch marks derived from electroglottograph (EGG) signals.
Testing involved both artificial reverberation (simulated white noise decay) and real-world Room Impulse Responses (RIRs) measured in a "SMART room," where the SMD was varied from 30 cm to 400 cm.
Results and Discussion
The Smearing Effect and Harmonic Proximity
Reverberation degrades speech signals by "smearing" their harmonic structure. In the time-frequency domain, the echoes of previous harmonics overlap with current ones, obscuring the clear periodic patterns required for $F_0$ estimation.
A key insight of this study is the "Harmonic Proximity Problem." Male voices typically possess lower fundamental frequencies, resulting in harmonics that are packed more densely in the spectrum. Consequently, even moderate reverberation causes these harmonics to overlap. In contrast, female voices, with their wider harmonic spacing, exhibit natural resilience to this spectral blurring.
Impact of Speaker-to-Microphone Distance (SMD)
The experiments revealed a stark contrast between near-field and far-field performance. In the near field (e.g., 30 cm), the direct sound is dominant, and $T_{60}$ has a negligible impact on accuracy. However, as the SMD increases, the Direct-to-Reverberation Ratio (DRR) decreases, and the specific acoustics of the room begin to dictate the error rate.
As shown in the performance metrics, female $F_0$ estimation remains remarkably stable across all tested distances. Conversely, male $F_0$ estimation accuracy (Fine Correction Rate) drops precipitously as the speaker moves into the far field. This confirms that $T_{60}$ alone cannot predict algorithm failure; the physical position of the speaker is a decisive factor.
Conclusion
This study demonstrates that reverberation-induced pitch tracking errors are highly dependent on the interaction between room geometry (SMD) and speaker gender. The findings suggest that future "reverb-aware" pitch trackers should account for frequency-dependent decay and the specific spectral characteristics of the speaker. Integrating these insights into dereverberation front-ends could significantly improve the robustness of ASR systems in challenging acoustic environments.
Summary updated