Natural speech statistics shift phoneme categorization

Abstract

All perception takes place in context. Recognition of a given speech sound is influenced by the acoustic properties of surrounding sounds. When the spectral composition of earlier (context) sounds (e.g., more energy at lower first formant [F 1 ] frequencies) differs from that of a later (target) sound (e.g., vowel with intermediate F 1), the auditory system magnifies this difference, biasing target categorization (e.g., towards higher-F 1 /ɛ/). Historically, these studies used filters to force context sounds to possess desired spectral compositions. This approach is agnostic to the natural signal statistics of speech (inherent spectral compositions without any additional manipulations). The auditory system is thought to be attuned to such stimulus statistics, but this has gone untested. Here, vowel categorization was measured following unfiltered (already possessing the desired spectral composition) or filtered sentences (to match spectral characteristics of unfiltered sentences). Vowel categorization was biased in both cases, with larger biases as the spectral prominences in context sentences increased. This confirms sensitivity to natural signal statistics, extending spectral context effects in speech perception to more naturalistic listening conditions. Importantly, categorization biases were smaller and more variable following unfiltered sentences, raising important questions about how faithfully experiments using filtered contexts model everyday speech perception.

Natural music context biases musical instrument categorization

“…Despite being matched to filtered contexts in duration (Experiment 1) and long-term average spectra (Experiments 1 and 2), unfiltered contexts produced smaller (if any) SCEs than filtered contexts in each experiment. This finding parallels Stilp and Assgari (2019), where unfiltered context sentences produced smaller and more variable SCEs in vowel categorization than filtered renditions of a single context sentence. In that study, each unfiltered block presented two different sentences sometimes spoken by two different talkers, similar to two different musical passages played by two different musical instruments (and musicians) presented here.…”

“…contribute to perception in more variable (and naturalistic) listening conditions. Recently, Stilp and Assgari (2019) observed SCEs in vowel categorization following highly controlled filtered sentences (as in previous studies) as well as sentences that naturally possessed the desired spectral properties without filtering, significantly enhancing the ecological validity of SCEs. Additionally, SCEs produced by unfiltered sentences were smaller than SCEs produced by filtered sentences, shedding considerable light on the precise degree to which these effects shape everyday perception.…”

“…In conclusion, the present experiments tested the influence of filtered or unfiltered musical contexts on musical instrument categorization. Following Stilp and Assgari (2019), this experimental paradigm deliberately sacrificed considerable acoustic control in order to test naturalistic (unfiltered) contexts that were more acoustically variable (as is commonly the case in everyday perception). Unfiltered musical passages biased perception of musical instruments (Experiment 2), but to a lesser degree than filtered contexts.…”

“…The horn context was a 3,032-ms excerpt of Sarah Willis of the Berlin Philharmonic playing French horn in the seldom-used unstopped playing style (i.e., hand not placed in the horn's bell to alter its spectrum, in order to match the unstopped horn target stimuli). Different musical selections were selected in order to model the high acoustic variability of everyday perception, and to parallel Stilp and Assgari (2019), who presented two different sentences in their unfiltered conditions. One-second excerpts of these stimuli were presented as contexts in Experiment 1, and were presented in their entireties in Experiment 2 (see Fig.…”

Acoustic context effects in speech perception

“…Later, Stilp () tested this question by presenting sentence contexts before /da/‐/ga/ targets. He measured the inherent balance of spectral energy across two frequency regions in sentences (low F 3 region: 1,700–2,700 Hz, high F 3 region: 2,700–3,700 Hz) using mean spectral differences (MSDs; Stilp & Assgari, ). MSDs were measured in two different temporal windows of the context sentences: the last 500 ms of the sentence (the Late window) and everything preceding the last 500 ms (the Early window).…”