Quantifying Talk-Time in Singing Instruction: Evaluation of an Automated Method

Introduction: Quantifying teacher and student talk-time during music lessons is crucial for effective voice instruction. Prior studies analyzing teacher talk-time relied on manual coding of recorded lessons (Burwell, 2018; Carey & Grant, 2014; Daniel 2006; Davies, 2011;). Such coding and analysis are error-prone, time-intensive, and hold weak statistical power, leading to a limited sample size. For this reason, speaking times need to be analyzed in an automated way. Addressing this, a neck-worn device that recorded audio via Zoom and provided automatic transcriptions was tested in a pilot study presented at the Voice Foundation’s 2023 meeting. Automated analysis of these transcriptions using a proprietary Python script showed promise in overcoming limitations, but revealed channel crosstalk issues introduced by varying teacher-student distance. Here, we address this issue by measuring the accuracy of automated speaking time detection while systematically varying teacher-student distance.

Methods: A sample “live” voice lesson was recorded that isolated speech from the singer and teacher into separate audio tracks. This audio was manually coded for singer-teacher talk time and speech content to establish a ground truth. In the experimental setup, this sample audio was played back through acoustic manikins with anatomically modeled vocal tracts to simulate the teacher-student acoustic interaction in a controlled laboratory setting. Manikin distance was controlled in six graduated half-meter steps, and the sample lesson audio was recorded at each distance. Zoom transcripts of the six simulated lessons were collected and automatically coded using the proprietary Python script. Data were analyzed with Bland–Altman plots (Ranganathan et al., 2017) to estimate agreement between the proposed automated method and manually coded data.

Results and Discussion:
Preliminary results suggest that this study may help to quantify the limitations of the automated method; it is expected that the findings of the study may drive more efficient research processes in the future. It is hoped that the result may increase our understanding of teachers’ pedagogical practice.