Background/Objectives
Voice signals can be classified into four types to assist in selecting a relevant analysis method for a specific signal type. The process of signal typing relies on visual spectrographic analysis, often leading to inconsistent results due to its subjective nature. To enhance the reliability of signal typing, this study aimed to investigate spectrographic characteristics and acoustic features associated with Types 1-4 signals, while examining the effectiveness of specific acoustic measures in differentiating these types and subcategories of spectrographic features.
Methods
This study analyzed sustained vowel and connected-speech samples from an existing dataset comprising control (n = 106), healthy-experimental (n = 12), and voice-disordered participants (n = 125). Spectrograms were generated, signal-typed, and evaluated on harmonics, subharmonics, modulations, and noise levels by three raters. Cepstral and spectral parameters, including Cepstral Peak Prominence (CPP), Low/High Spectral Ratio (LH), and Cepstral/Spectral Index of Dysphonia (CSID), were measured using the Analysis of Dysphonia in Speech and Voice (ADSV). Frequency-based measures including speaking Fundamental Frequency (F0), Jitter, Shimmer, Harmonics-to-noise Ratio (HNR), and F0 Standard Deviation (F0SD), were extracted from Types 1 and 2 samples using Praat. Discriminant function analysis and logistic regression were used to determine the discrimination values of these acoustic measures in identifying signal types and spectrographic subcategories, respectively.
Results
Results indicated a progressive reduction in harmonic visibility in Types 3 and 4 signals, making subharmonic and modulation analysis impossible. Types 2 to 4 exhibited increasing spectral noise, reflecting diminished periodicity. CPP and CSID achieved higher classification accuracy (>75%) than the other measures, yet no individual measure consistently differentiated all signal types. Frequency-based measures significantly discriminated spectrographic subcategories, while spectral measures showed the strongest signal type discrimination accuracy.
Conclusions
This research supports a refined signal-typing framework by establishing specific spectrographic characteristics for more reliable visual analysis. Frequency-based measures provided were found to inform spectrographic analysis, while spectral measures demonstrated stronger discriminatory accuracy between signal types. These findings suggest that integrating these spectral acoustic parameters into machine-learning algorithms could enhance the objectivity and reliability of automated signal typing in clinical practice.