Duration of ambulatory monitoring needed to accurately estimate voice use

Daryush D. Mehta, Rebecca Woodbury Listfield, Harold A. Cheyne Ii, James T. Heaton, Shengran W. Feng, Matías Zañartu, Robert E. HillmanInterspeech 2012

TL;DR

This paper investigates how long individuals need to be monitored using an accelerometer-based device to accurately estimate their voice use. They found that while fundamental frequency and sound pressure level stabilize relatively quickly, phonation time, cycle dose, and distance dose require longer monitoring periods (around 26 hours) for reliable estimation. This work provides crucial benchmarks for designing future voice monitoring studies and clinical applications.

What Problem Does This Paper Address?

Voice disorders are often linked to how individuals use their voices throughout the day. Clinicians currently rely on subjective self-reports to assess voice use, which can be unreliable due to habituated vocal behaviors often occurring below conscious awareness. This lack of objective data hinders our understanding of the role of vocal loading in voice disorders and the establishment of normal voice use parameters. To address this, wearable devices like the Ambulatory Phonation Monitor (APM) have been developed to objectively track voice use measures such as phonation time, fundamental frequency, and sound intensity. However, there's a lack of established guidelines for how long these devices need to be worn to obtain accurate and representative estimates of voice use. This paper aims to fill that gap by providing initial benchmarks for the duration of ambulatory monitoring required for robust voice use parameter estimation.

Methods

The study employed the Ambulatory Phonation Monitor (APM), a neck-mounted accelerometer that unobtrusively senses phonation-induced vibrations. The APM processes this signal in real-time to estimate fundamental frequency, sound pressure level, and phonation time. For calibration, speech was recorded simultaneously with a handheld microphone positioned near the subject's mouth.

The core of the data analysis involved a cumulative averaging technique. For each hour of monitoring, the APM provided estimates for various parameters. These hourly estimates were then used to calculate cumulative averages. For instance, the cumulative average for the second hour is the average of the first and second hour's data. A "complete sample" was defined as 40 hours of monitoring, serving as the "true value" for each parameter. The researchers then calculated the percent error for each parameter at each hour by comparing the cumulative average up to that hour (estimated value) with the 40-hour cumulative average (true value). This allowed them to track how estimation accuracy improved over time.

Datasets

The study involved three groups of subjects, each with six participants:

  • Patients with voice disorders: Recruited from a voice center, exhibiting mild to severe dysphonia and requiring substantial voice use in their occupations.
  • Control subjects: Matched to the patient group in terms of sex, approximate age, and occupation.
  • Low voice users: Individuals with occupations not requiring significant daily talking, recruited from research staff.

All subjects underwent laryngeal endoscopic examination to confirm the presence or absence of pathology. Data was collected over five work days per subject, yielding 30-70 hours of monitoring data per individual. For the analysis of error curves, only subjects who were monitored for a minimum of 40 hours were included (3 patients, 4 controls, and 3 low voice users), totaling 10 subjects.

Key Insights and Ideas

  • Variability is Key: The required monitoring duration for accurate estimation is directly tied to the inherent variability of the voice use parameter being measured. Parameters with lower variability stabilize faster.
  • Fundamental Frequency and SPL are Quick to Stabilize: These acoustic measures appear to reach a stable estimate relatively quickly, suggesting that shorter monitoring periods might suffice for their assessment.
  • Phonation Time, Cycle Dose, and Distance Dose Need More Time: These measures, which reflect the cumulative effort and duration of vocal fold vibration, are more sensitive to daily fluctuations and require significantly longer monitoring to achieve reliable estimates.
  • Individual Differences Matter: While average trends are informative, the study acknowledges that individual subject behavior can lead to significant variations in error, even after prolonged monitoring. This highlights the need for personalized monitoring durations in some cases.
  • Smartphone Integration is the Future: The paper hints at the development of a new system utilizing smartphones, which promises improved user comfort, compliance, and the potential for more advanced analysis beyond current APM capabilities.

Key Results

  • Fundamental Frequency (f0) Error: Average f0 error decreased to approximately 1% after 12 hours of monitoring, with a worst-case scenario showing less than 5% error.
  • Sound Pressure Level (SPL) Error: Average SPL error reached about 5% after just one hour and approximately 1% after 20 hours of monitoring.
  • Phonation Time Error: An average error of approximately 5% was achieved after 26 hours, with most error variation decreasing thereafter. However, one individual showed errors up to 18% even after prolonged monitoring.
  • Cycle Dose and Distance Dose Error: Average errors for these parameters remained above 10% until the 26-hour mark. Individual errors could reach up to 19% for cycle dose and 26% for distance dose after prolonged monitoring.

Limitations and Future Work

Limitations: * Small Sample Size: The authors acknowledge that the current data is based on a small sample size (10 subjects for the error curve analysis). * Speaker-to-Speaker Variation: They also note that day-to-day variations in phonatory behavior among speakers could confound the notion of overall parameter averages.

Future Work: * Larger Subject Sample: Gathering data from a larger and more diverse group of subjects is crucial for more robust conclusions. * Next-Generation Monitor: Implementing a next-generation ambulatory voice monitor is planned. This new system, utilizing a smartphone as the data acquisition device, aims to improve user experience and enable the characterization of individual vocal behavior, including vocal recovery times. * Advanced Analysis: The raw waveform collection capability of the new system will allow for the investigation of alternative analysis algorithms, such as deriving glottal airflow-based measures and employing pattern recognition techniques to identify potentially abusive vocal patterns.

Summary updated

BibTeX
@inproceedings{mehta2012duration,
  author = {Daryush D. Mehta and Rebecca Woodbury Listfield and Harold A. Cheyne II and James T. Heaton and Shengran W. Feng and Matías Zañartu and Robert E. Hillman},
  title = {{Duration of ambulatory monitoring needed to accurately estimate voice use}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure