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

Felix Schaeffler, Stephen Jannetts, Janet BeckInterspeech 2019

Sure, here's a blog-style summary of the paper "Reliability of Clinical Voice Parameters Captured with Smartphones — Measurements of Added Noise and Spectral Tilt":

TL;DR

This paper investigates the reliability of smartphone voice recordings for clinical analysis, focusing on measures of added noise and spectral tilt. It finds that while some parameters like HNR and GNE show promising reliability, others like spectral tilt and H1-H2 exhibit significant random errors, suggesting that careful, parameter-specific assessment is crucial before widespread clinical adoption of smartphone voice analysis.

What Problem Does This Paper Address?

Smartphones are increasingly being explored as tools for clinical voice analysis due to their ubiquity and advanced sensors. However, the reliability of voice parameters extracted from smartphone recordings compared to professional studio equipment is still a significant concern. Previous studies have yielded mixed results, with some suggesting sufficient reliability while others highlight potential issues. This paper specifically addresses the gap in understanding the reliability of measures related to added noise and spectral tilt, which are crucial for assessing voice quality and health, by systematically evaluating systematic and random errors across different smartphone models and comparing them to a studio microphone.

Methods

The study builds upon a previous investigation by analyzing six acoustic parameters: Harmonics-to-noise ratio (HNR), Glottal noise excitation ratio (GNE), and four spectral tilt measures (Slope, Tilt, H1-H2, and H1-H2). Acoustic analysis was performed using the Praat software. The core methodology involved a Bland-Altman analysis to quantify both systematic error (bias) and random error (critical difference) for each parameter and smartphone. The critical difference was then related to the total range of values observed with the studio microphone to assess its practical relevance, with a preliminary guideline suggesting that random errors not exceeding 10% of the total range might be acceptable.

Datasets

The study utilized data collected from 12 women and 10 men who recorded sustained [a] vowel sounds and a short passage. These recordings were captured simultaneously using one studio microphone and four different smartphones: a Samsung Galaxy S8+, an iPhone 6s, and an iPhone 7. The passage data was prioritized for analysis due to its generally higher reliability in previous studies.

Key Insights and Ideas

  • Parameter-Specific Reliability is Key: The study strongly emphasizes that smartphone reliability for voice analysis is not a one-size-fits-all situation. Different acoustic parameters exhibit vastly different levels of reliability when captured by smartphones.
  • Added Noise Measures Show Promise: Parameters like HNR and GNE demonstrated relatively good reliability, with critical differences often falling within acceptable ranges, suggesting their potential for clinical use with smartphones.
  • Spectral Tilt is Problematic: Measures of spectral tilt, particularly the uncorrected H1-H2, showed substantial random errors, raising concerns about their suitability for precise clinical assessment using smartphones.
  • Random Error Distribution Matters: For some parameters, especially H1-H2, the Bland-Altman plots indicated that the random error might not be uniformly distributed across the range of values, which warrants further investigation.
  • Context is Crucial for Error Interpretation: The authors highlight that the relevance of systematic and random errors depends on the clinical context. Systematic errors might be less critical for within-speaker comparisons but crucial for inter-device comparisons or when calibrating against studio equipment. Random errors, however, always impact the interpretation of change.

Key Results

  • HNR: Fulfilled the 10% criterion for random error on the Samsung devices (8.6% and 9.2%) and was close for iPhones (10.5% and 11.0%).
  • GNE: Showed critical differences between 13.9% and 20.8% of the range across all devices.
  • Slope: Had critical differences between 13.5% and 16.9% of the range for most devices.
  • Tilt: Exhibited high relative random errors, with critical differences for Samsung devices at 47.4-51.3% and for iPhones at 21.6-27.8% of the range.
  • H1-H2: Showed critical differences in the range of 22.1-36% for the uncorrected measure.
  • H1-H2: Had critical differences ranging from 16.2% to 54.5%, with Samsung devices showing higher critical differences than Apple devices.

Limitations and Future Work

The authors acknowledge that the preliminary guideline of 10% random error is somewhat ad-hoc and that further investigation into the practical relevance of these errors is needed. They also point out that the study focused on typical voices, and the extension to pathological voices is speculative and requires further study. Future work should explore how calibration can mitigate systematic errors and how environmental factors like ambient noise and room configuration impact measurements. Additionally, more general studies on typical and clinical ranges of acoustic parameters would support the overall validity and reliability of clinical acoustic analysis using smartphones.

Summary updated

BibTeX
@inproceedings{schaeffler2019reliability,
  author = {Felix Schaeffler and Stephen Jannetts and Janet Beck},
  title = {{Reliability of Clinical Voice Parameters Captured with Smartphones — Measurements of Added Noise and Spectral Tilt}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure