Speech data collection at a distance: Comparing the reliability of acoustic cues across homemade recordings

Abstract

Speech production data collection has been significantly impacted by COVID-19 restrictions. Sound-treated recording spaces and high-quality recording devices are inaccessible, and face-to-face interactions are limited. We investigated alternative recording methods that produce data suitable for phonetic analysis, and are accessible to people in their homes. We examined simultaneous recordings of pure tones at seven frequencies (50 Hz, every 100 Hz between 100 Hz and 600 Hz), and three repetitions of the primary cardinal vowels elicited from five trained speakers. Recordings were made using the ZOOM meeting application and non-lossy format smartphone applications (Awesome Voice Recorder, Recorder), comparing these with Zoom H6N reference recordings. F0, F1-5, and duration based on manual segmentation were measured. F0 is highly correlated between the three devices for vowels and tones. Lower formants are also significantly correlated though not as robustly. The upper formants showed more variation as reported in the literature. Both phone and ZOOM performed better for vowels than tones. Phone segmentation generated reliable duration values differing from H6N segmentation by ∼18 ms. However, irregular waveforms and filtering algorithm artefacts caused considerable differences for ZOOM (∼119 ms). Our preliminary study suggests phone recordings are a viable option for some phonetic studies (e.g., prosody). Future analysis of natural speech data will prove insightful.