Assessing voice health using smartphones: bias and random error of acoustic voice parameters captured by different smartphone types

Abstract

Background Occupational voice problems constitute a serious public health issue with substantial financial and human consequences for society. Modern mobile technologies such as smartphones have the potential to enhance approaches to prevention and management of voice problems. This paper addresses an important aspect of smartphone‐assisted voice care: the reliability of smartphone‐based acoustic analysis for voice health state monitoring. Aim To assess the reliability of acoustic parameter extraction for a range of commonly used smartphones by comparison with studio recording equipment. Methods & Procedures Twenty‐two vocally healthy speakers (12 female, 10 male) were recorded producing sustained vowels and connected speech under studio conditions using a high‐quality studio microphone and an array of smartphones. For both types of utterance, Bland–Altman analysis was used to assess overall reliability for mean F0, cepstral peak prominence (CPPS), Jitter (RAP) and Shimmer %. Outcomes & Results Analysis of the systematic and random error indicated significant bias for CPPS across both sustained vowels and passage reading. Analysis of the random error of the devices indicated that that mean F0 and CPPS showed acceptable random error size, while jitter and shimmer random error was judged as problematic. Conclusions & Implications Confidence in the feasibility of smartphone‐based voice assessment is increased by the experimental finding of high levels of reliability for some clinically relevant acoustic parameters, while the use of other parameters is discouraged. We also challenge the practice of using statistical tests (e.g., t‐tests) for measurement reliability assessment.

Reliability of Clinical Voice Parameters Captured with Smartphones — Measurements of Added Noise and Spectral Tilt

“…The study by [7] therefore suggests that smartphone reliability for acoustic voice assessment very much depends on the parameter under question. The study presented here is a continuation study of [7], extending the analysis to the quantification of added noise and glottal pulse efficiency as reflected in the spectral tilt of the voice spectrum.…”

“…Smartphones are increasingly used to collect speech data for acoustic analysis, and several authors have suggested that smartphones have sufficient reliability for voice analysis with a clinical purpose [1]- [6]. [7] took a less optimistic stance on this and argued that previous analyses partly rested on false assumptions about the nature of the problem. For example, some studies have used non-significance between a smartphone recording and a reference recording as evidence that smartphone measurement are reliable.…”

“…For example, some studies have used non-significance between a smartphone recording and a reference recording as evidence that smartphone measurement are reliable. [7] have argued that this approach can only ever assess systematic error and that the random error of a measurement is completely ignored with this approach, or even worse, a high random error supports a non-significant result and thus a 'positive' conclusion. If both systematic and random error are quantified, then some acoustic parameters show worrying levels of random error.…”

“…If both systematic and random error are quantified, then some acoustic parameters show worrying levels of random error. [7] investigated four acoustic parameters with high relevance for clinical voice analysis, mean F0, smoothed Cepstral Peak Prominence (CPPS), jitter (RAP) and shimmer %, across four popular smartphone devices (two Samsung and two Apple models) and for two different types of voice material, i.e. sustained vowels and the reading of a passage.…”

“…The random error was also below ±1 dB, but assessing the relevance of this amount of random error is more difficult than for F0. [7] suggested relating the random error to the range of values observed in a sample and to clinical thresholds, if available. In the study sample, the random error of CPPS corresponded to about 10% of the total range of values.…”