Auditory Reverse Correlation on a Phoneme-Discrimination Task: Assessing the Effect of Different Types of Background Noise
Résumé
Background:
A long-lasted goal in psycholinguistics is to identify the acoustic cues underlying phonetic percepts. The reverse correlation method offers an agnostic approach to explore internal representations of phonemes by relating listeners’ responses in a phoneme-in-noise task with the exact time-frequency representation of the tested noises. The outcome is called auditory classification images (ACIs): A fine-grained time-frequency map of the acoustic cues listeners relied upon. Here, we focus on the effect of the statistics of the background noises with respect to the efficiency and robustness of the method. We used three different noise types that have a flat long-term spectrum, but differ in the amount of temporal envelope fluctuations: (1) White noise, (2) white noise low-pass filtered in the modulation power spectrum (MPS) domain, and (3) bump noise.
Methods:
We conducted a consonant-in-noise discrimination task using the words /aba/ and /ada/. The speech samples were uttered by a male speaker taken from the French Logatome speech corpus. Each participant performed a total of 4000 /aba/-/ada/ categorizations in each of the three noise conditions. During the task, the signal-to-noise ratio (SNR) was varied to target a score of 70.7%. Subsequently, ACIs were derived using the reverse correlation method, revealing the time-frequency regions where the noises systematically affected the participants’ discrimination. Further, we computed ACIs using subsets of trials to evaluate the fidelity and speed of convergence of these "partial ACIs" with respect to the ACI obtained using all trials.
Results:
For both participants and for the three test noise types, we succeeded to find stable ACIs which, in line with previous evaluations that used another pair of /aba/-/ada/ utterances, reveal a high-weighting of the speech information between the first and second formants in the consonant-vowel transitions. Our analysis of partial ACIs shows that the three noise types required a different number of conducted trials to converge to a stable ACI, with a faster convergence for the MPS and bump noises compared with the use of white noises.
Our analyses suggest that the reverse correlation method applied to a consonant-in-noise discrimination task converges more quickly to a stable result, the ACI, when the background noises contain dominant components in the modulation frequency range between 0 and 40 Hz, which is the case for MPS and bump noises. The prominent envelope fluctuations in this range lead to more systematic confusion errors compared to white noise and, therefore, to higher prediction accuracy and more robust reverse correlation results.
Origine : Fichiers produits par l'(les) auteur(s)