(Don't) try this at home! The effects of recording devices and software on phonetic analysis

Abstract

Because of restrictions on in-person research due to COVID-19, researchers are now relying on remotely recorded data to a much greater extent than in the past. Given the change in methodology, it is important to know how remote recording might affect acoustic measurements, either because of different recording devices used by participants and consultants or because of the software used to make recordings. This study investigates audio signal fidelity across different inperson recording equipment and remote recording software when compared to solid-state digital audio. We show that equipment choice and software can have a large effect on acoustic measurements, including those of frequency, duration, and noise. The issues do not just reflect decreased reliability; some measurements are systematically shifted in particular recording conditions. These results show the importance of carefully considering and documenting equipment choices, particularly for crosslinguistic or cross-speaker comparisons. We close with a framework for researchers to use in deciding what types of recording may be most appropriate.*

Abstracts of the 8th International Conference on Speech Motor Control Groningen, August 2022

“…Thus, STI changes in response to slow speech may not be solely driven by FORMANT MEASUREMENTS FROM REMOTELY-COLLECTED WAV AND M4A FILES ARE SIMILAR Karen V. Chenausky 1,2,3 , Helen Tager-Flusberg 3 , Michaela Flaherty 1 ,Jordan Green 11 MGH Institute of Health Professions, Boston, Massachusetts, USA 2 Harvard Medical School, Boston, Massachusetts, USA 3 Boston University, Boston, Massachusetts, USA One of the challenges of remote speech data collection is that apps using lossy compression audio file formats are more common than lossless file formats. While previous work (Sanker et al, 2021) suggests that there is no significant effect of file format on mean F0, F1, or F2 from adult speech, no similar comparisons have been performed for children's speech, especially children who may have speech disorders. To determine whether F0 and formant frequency measurements for (possibly disordered) children's speech differed according to file format, we compared measurements made across WAV and M4A file types.…”

Medium-shifting and intraspeaker variation in conversational interviews

“…Because vowel normalization procedures commonly make use of the third formant in vocal tract length calculations (Johnson, 2020), these distortions pose an issue for traditional methods of speaker normalization. Sanker et al (2021), however, did not find any overall differences in formant measurements between Zoom and handheld recorders, though the authors note that the lack of significance is not indicative of reliability, but rather that different vowels are impacted in different ways.…”

“…1.The solid-state recording device used as a baseline in these studies is the Zoom Handy Recorder (model H6 in Zhang et al, 2021; H4n in Sanker et al, 2021), which is popular among sociolinguists and phoneticians working in the field. This is not to be confused with the unrelated Zoom video conferencing program.…”

“…Zhang, Jepson, Lohfink, and Arvaniti (2021) found that audio recordings from Zoom showed lower values for F1, F2, and F3 in vowels, and unexpected fluctuations in intensity, compared to analogous recordings made on a solid-state digital recording device with no file compression. Sanker, Babinski, Burns, Evans, Johns, Kim, Smith, Weber, and Bowern (2021) did not find any consistent overall differences in vowel formant frequency attributable to Zoom, though they note that there may still be meaningful effects that are vowel-specific; at the same time, the authors identify effects related to duration (shorter consonants and longer vowels, as determined by a forced aligner) and a higher signal-to-noise ratio, possibly stemming from Zoom's proprietary background noise reduction algorithm 1 . These findings should inspire caution when analyzing and interpreting raw acoustic measurements from recorded video calls, especially when relying on recordings from multiple mixed media in a single study.…”

The effectiveness of embodied prosodic training in L2 accentedness and vowel accuracy

“…Sanker et al (2021) provided suggestions for online experiments in linguistic fieldwork, phonetics, and other studies involving speech data collection. Our methodology conformed to Sanker et al's (2021) main recommendations in that (a) the participants did the recordings on their own devices and directly uploaded them onto online storage services (Google Drive) to avoid further compression; (b) the recording platform was the same across participants and sessions, and (c) our sample size was reasonable for group-level comparisons which avoided individual-level analyses. However, it was not possible to (a) make uniform participants' computer setups, (b) install external microphones for each of them, or (c) generate non-compressed recordings from the free online recording service.…”