Face masks and the perception of fricatives

During the COVID-19 pandemic, face masks became compulsory in many countries, and many people felt that masked speech was harder to understand. A mask can affect speech perception in two ways: it filters the sound, and it hides the lips and teeth. This study asked which of the two matters more for four English fricatives: /f/ and /θ/, whose articulation is visible, and /s/ and /ʃ/, whose articulation is largely hidden.

Task

Listeners watched short videos of single words and pressed the button for the fricative they heard. Twenty words formed minimal sets in onset position (fie, thigh, sigh, shy), twenty formed minimal pairs in coda position (deaf, death), and forty had no competitor (fact, shame).

Results: accuracy and reaction times

The mask lowered accuracy only for the fricatives with visible articulation: most of all for /θ/ (z = −10.01, p < .0001), then for /f/ (z = −6.84, p < .0001). /s/ and /ʃ/ were recognised almost perfectly in both conditions; in the mixed-effects model, the interaction of condition and fricative was significant (χ² = 53.51, p < .0001). Reaction times were longer with the mask (χ² = 9.14, p < .005), but by a similar amount for all four fricatives: the interaction was not significant. Sibilants were identified faster than non-sibilants, and the device (computer or mobile) also affected reaction times (χ² = 17.52, p < .0001).

Results: confusions

With the mask, /θ/ was heard as /f/ in about 10% of trials, compared with 2% without it. The confusion also runs the other way: /f/ was heard as /θ/ in 13% of trials without a mask and in 18% with one, which is why /f/ had the lowest accuracy in both conditions. The paper relates this to the ambiguously encoded (labio)dental contrast in English.

Results: acoustics or visual cues?

The mask also changed the sound. It lowered the centre of gravity most in /f/ (t(76) = −8.91, p < .0001) and /θ/ (t(76) = −5.18, p < .0001), but not in /s/ or /ʃ/, and it lowered intensity most in /s/ (t(76) = −9.68, p < .0001), followed by /f/, /ʃ/ and /θ/. Yet /s/, the fricative with the largest drop in intensity, did not become harder to recognise. The paper concludes that the effect of the mask comes mainly from the missing visual cues to the (labio)dental fricatives, together with their low acoustic energy. Acoustic filtering played a smaller role.

The acoustic measures come from the stimuli: each word was recorded once in each condition. The correlations are descriptive, with 20 words per fricative.

Results: lexical frequency

Frequent words were recognised more accurately (χ² = 8.00, p < .005) and faster (χ² = 7.91, p < .005), but this was equally true with and without the mask. Lexical competition (minimal pairs) did not significantly affect accuracy or reaction times. In the plot below, each word is shown with its accuracy in both conditions; the search box finds a word, and selecting a word shows its details. The lines are simple least-squares fits of the raw proportions, for illustration only.

Results: listeners

Participants from England and Scotland were coded as speakers of varieties with th-fronting, the merger of /θ/ with /f/ (as in fink for think). The mask did not affect them differently from the other listeners, neither in accuracy (χ² = 0.57, n.s.) nor in reaction times, although they were slower overall in responding to /f/ and /θ/. Each point in the plot below is one listener; points below the diagonal are listeners who were less accurate with the mask.

Reference

Canzi, M., & Rathcke, T. (2023). Unmasking the truf: Impact of community masks on the perception of voiceless fricatives in English. In Proceedings of the 20th International Congress of Phonetic Sciences, Prague.

Analyses in R with mixed-effects regression (lme4, emmeans); the statistics quoted in the text come from the paper’s models. The figures on this page show the study’s raw data: accuracy is the proportion of correct responses per listener, averaged over listeners; reaction times are geometric means of correct responses between 150 and 5000 ms.