Abstract
This corrects the article on p. 1 in vol. 4, PMID:
26835449.
Introduction
This corrigendum addresses a visual scaling error in the original publication, "Voice Pathology Detection Using Modulation Spectrum-Optimized Metrics." The correction focuses on Figure 5, ensuring the x-axis accurately reflects the experimental range of frame lengths discussed in the study's text.
The Importance of Frame Length in Dysphonia Detection
Automated voice pathology detection relies on digital signal processing to identify disorders from speech recordings. A pivotal parameter in this analysis is the frame length—the duration of the temporal window analyzed. Selecting an optimal frame length is a balancing act: windows that are too short sacrifice frequency resolution, while excessively long windows lose the temporal resolution necessary to track rapid vocal changes.
The original study investigated how varying frame lengths (from 20 ms to 200 ms) affects the efficiency of metrics derived from the modulation spectrum. Unlike standard frequency analysis, modulation spectrum analysis captures the micro-instabilities (shimmer and jitter) characteristic of pathological voices by examining how amplitude and frequency change over time.
Corrected Experimental Results
The primary correction involves Figure 5, which illustrates the detection efficiency of various metrics across different frame lengths. In the original publication, the x-axis was truncated at 120 ms, despite the text discussing results up to 200 ms. The corrected figures now span the full range, confirming the trends observed in the data.
Metric Sensitivity and the Plateau Effect
The data reveals a distinct "plateau effect" across most metrics. Efficiency typically increases sharply as the frame length grows from 20 ms to approximately 80–100 ms. Beyond this point, further increases in the analysis window yield diminishing returns in detection accuracy.
Different metrics exhibit varying degrees of sensitivity to the analysis window. For instance, the RALA and MSP95 (Modulation Spectrum Percentile) metrics show significant improvement in the 20–60 ms range. This suggests these specific metrics require a minimum temporal window to effectively capture the modulation characteristics of a pathological voice.
Database Disparity
The results also highlight a consistent performance gap between the two benchmark datasets used: the MEEI (Massachusetts Eye and Ear Infirmary) database and the PdA (Príncipe de Asturias Hospital) database. Performance is generally higher on the MEEI data, which is often more controlled than the diverse, real-world recordings found in the PdA database. This disparity underscores the ongoing challenge of developing robust models that generalize across different recording environments and patient populations.
Conclusion
While the visual representation of the data was initially incomplete, the underlying quantitative analysis and the resulting scientific conclusions remain unchanged. The corrected Figure 5 reinforces the study's findings regarding the optimal frame length for modulation spectrum-based voice pathology detection.
Summary updated