Abstract
INTRODUCTION: The Bridge2AI-Voice. Consortium is developing affordable and accessible voice data to assist in the identification of vocal biomarkers of disease in adults and children. Initial experiments were designed to establish voice recording procedures to be used in research labs and clinical settings, as well as in quiet environments outside of the clinic. The focus has been on isolated vowel productions, which provide a vocal signal that is representative of the biomechanics of the larynx within a static vocal tract. The current experiment considers the impact of sentence productions on the measurement of several acoustic parameters. METHODS: Voice recordings from 24 individuals representing a wide range of typical and disordered voices were analyzed. Two CAPE-V sentences were recorded via a head-and-torso model using (1) a research quality, clinical standard microphone/preamplifier/audio interface and (2) smartphones and tablets using their internal microphones and an attached external headset microphone. Mouth-to-microphone distances and environmental noise levels were controlled. Measures of fundamental frequency (F(0)) and spectral and cepstral measures of voice quality valid for use in sentence contexts were analyzed across recording conditions. RESULTS: Cepstral peak prominence (CPP) values were sensitive to microphone type, noise, and sentence type conditions. Nevertheless, strong linear relationships were observed across recording methods compared to the clinical standard. Measures of F(0) obtained using autocorrelation correlated strongly across recording methods, whereas F(0) measures obtained from the CPP (CPP F(0)) were highly variable and poorly correlated across recording methods and noise conditions. The L/H ratio (a measure of spectral tilt) was significantly affected by recording condition but not background noise, and measures of L/H ratio were also observed to correlate strongly across recording methods and noise conditions. DISCUSSION: Current findings revealed that different recording methods can produce significantly different acoustic measures of voice with sentence-level materials. Since microphone characteristics (e.g., frequency response; use of noise cancellation), mouth-to-microphone distances, and background noise conditions can have significant effects on spectral and cepstral assessment of voice, it is essential that recording methods and conditions are explicitly described when designing voice data collection projects and comparing datasets as it may have an impact on voice analysis. Future investigations should evaluate consistency of results among multiple examples of the same device.