Influences of selective adaptation on perception of audiovisual speech

Synthetic faces generated with the facial action coding system or deep neural networks improve speech-in-noise perception, but not as much as real faces

The prevalence of synthetic talking faces in both commercial and academic environments is increasing as the technology to generate them grows more powerful and available. While it has long been known that seeing the face of the talker improves human perception of speech-in-noise, recent studies have shown that synthetic talking faces generated by deep neural networks (DNNs) are also able to improve human perception of speech-in-noise. However, in previous studies the benefit provided by DNN synthetic faces was only about half that of real human talkers. We sought to determine whether synthetic talking faces generated by an alternative method would provide a greater perceptual benefit. The facial action coding system (FACS) is a comprehensive system for measuring visually discernible facial movements. Because the action units that comprise FACS are linked to specific muscle groups, synthetic talking faces generated by FACS might have greater verisimilitude than DNN synthetic faces which do not reference an explicit model of the facial musculature. We tested the ability of human observers to identity speech-in-noise accompanied by a blank screen; the real face of the talker; and synthetic talking faces generated either by DNN or FACS. We replicated previous findings of a large benefit for seeing the face of a real talker for speech-in-noise perception and a smaller benefit for DNN synthetic faces. FACS faces also improved perception, but only to the same degree as DNN faces. Analysis at the phoneme level showed that the performance of DNN and FACS faces was particularly poor for phonemes that involve interactions between the teeth and lips, such as /f/, /v/, and /th/. Inspection of single video frames revealed that the characteristic visual features for these phonemes were weak or absent in synthetic faces. Modeling the real vs. synthetic difference showed that increasing the realism of a few phonemes could substantially increase the overall perceptual benefit of synthetic faces.

The Effect on Speech-in-Noise Perception of Real Faces and Synthetic Faces Generated with either Deep Neural Networks or the Facial Action Coding System

The prevalence of synthetic talking faces in both commercial and academic environments is increasing as the technology to generate them grows more powerful and available. While it has long been known that seeing the face of the talker improves human perception of speech-in-noise, recent studies have shown that synthetic talking faces generated by deep neural networks (DNNs) are also able to improve human perception of speech-in-noise. However, in previous studies the benefit provided by DNN synthetic faces was only about half that of real human talkers. We sought to determine whether synthetic talking faces generated by an alternative method would provide a greater perceptual benefit. The facial action coding system (FACS) is a comprehensive system for measuring visually discernible facial movements. Because the action units that comprise FACS are linked to specific muscle groups, synthetic talking faces generated by FACS might have greater verisimilitude than DNN synthetic faces which do not reference an explicit model of the facial musculature. We tested the ability of human observers to identity speech-in-noise accompanied by a blank screen; the real face of the talker; and synthetic talking face generated either by DNN or FACS. We replicated previous findings of a large benefit for seeing the face of a real talker for speech-in-noise perception and a smaller benefit for DNN synthetic faces. FACS faces also improved perception, but only to the same degree as DNN faces. Analysis at the phoneme level showed that the performance of DNN and FACS faces was particularly poor for phonemes that involve interactions between the teeth and lips, such as /f/, /v/, and /th/. Inspection of single video frames revealed that the characteristic visual features for these phonemes were weak or absent in synthetic faces. Modeling the real vs. synthetic difference showed that increasing the realism of a few phonemes could substantially increase the overall perceptual benefit of synthetic faces, providing a roadmap for improving communication in this rapidly developing domain.

Repetitive Exposure to Orofacial Somatosensory Inputs in Speech Perceptual Training Modulates Vowel Categorization in Speech Perception

Orofacial somatosensory inputs may play a role in the link between speech perception and production. Given the fact that speech motor learning, which involves paired auditory and somatosensory inputs, results in changes to speech perceptual representations, somatosensory inputs may also be involved in learning or adaptive processes of speech perception. Here we show that repetitive pairing of somatosensory inputs and sounds, such as occurs during speech production and motor learning, can also induce a change of speech perception. We examined whether the category boundary between /ε/ and /a/ was changed as a result of perceptual training with orofacial somatosensory inputs. The experiment consisted of three phases: Baseline, Training, and Aftereffect. In all phases, a vowel identification test was used to identify the perceptual boundary between /ε/ and /a/. In the Baseline and the Aftereffect phase, an adaptive method based on the maximum-likelihood procedure was applied to detect the category boundary using a small number of trials. In the Training phase, we used the method of constant stimuli in order to expose participants to stimulus variants which covered the range between /ε/ and /a/ evenly. In this phase, to mimic the sensory input that accompanies speech production and learning in an experimental group, somatosensory stimulation was applied in the upward direction when the stimulus sound was presented. A control group (CTL) followed the same training procedure in the absence of somatosensory stimulation. When we compared category boundaries prior to and following paired auditory-somatosensory training, the boundary for participants in the experimental group reliably changed in the direction of /ε/, indicating that the participants perceived /a/ more than /ε/ as a consequence of training. In contrast, the CTL did not show any change. Although a limited number of participants were tested, the perceptual shift was reduced and almost eliminated 1 week later. Our data suggest that repetitive exposure of somatosensory inputs in a task that simulates the sensory pairing which occurs during speech production, changes perceptual system and supports the idea that somatosensory inputs play a role in speech perceptual adaptation, probably contributing to the formation of sound representations for speech perception.

Acoustic-phonetic and auditory mechanisms of adaptation in the perception of sibilant fricatives

Listeners are highly proficient at adapting to contextual variation when perceiving speech. In the present study, we examined the effects of brief speech and nonspeech contexts on the perception of sibilant fricatives. We explored three theoretically motivated accounts of contextual adaptation, based on phonetic cue calibration, phonetic covariation, and auditory contrast. Under the cue calibration account, listeners adapt by estimating a talker-specific average for each phonetic cue or dimension; under the cue covariation account, listeners adapt by exploiting consistencies in how the realization of speech sounds varies across talkers; under the auditory contrast account, adaptation results from (partial) masking of spectral components that are shared by adjacent stimuli. The spectral center of gravity, a phonetic cue to fricative identity, was manipulated for several types of context sound: /z/-initial syllables, /v/-initial syllables, and white noise matched in long-term average spectrum (LTAS) to the /z/-initial stimuli. Listeners' perception of the /s/-/ʃ/ contrast was significantly influenced by /z/-initial syllables and LTAS-matched white noise stimuli, but not by /v/-initial syllables. No significant difference in adaptation was observed between exposure to /z/-initial syllables and matched white noise stimuli, and speech did not have a considerable advantage over noise when the two were presented consecutively within a context. The pattern of findings is most consistent with the auditory contrast account of short-term perceptual adaptation. The cue covariation account makes accurate predictions for speech contexts, but not for nonspeech contexts or for the absence of a speech-versus-nonspeech difference.

Attentional resources contribute to the perceptual learning of talker idiosyncrasies in audiovisual speech

To recognize audiovisual speech, listeners evaluate and combine information obtained from the auditory and visual modalities. Listeners also use information from one modality to adjust their phonetic categories to a talker's idiosyncrasy encountered in the other modality. In this study, we examined whether the outcome of this cross-modal recalibration relies on attentional resources. In a standard recalibration experiment in Experiment 1, participants heard an ambiguous sound, disambiguated by the accompanying visual speech as either /p/ or /t/. Participants' primary task was to attend to the audiovisual speech while either monitoring a tone sequence for a target tone or ignoring the tones. Listeners subsequently categorized the steps of an auditory /p/-/t/ continuum more often in line with their exposure. The aftereffect of phonetic recalibration was reduced, but not eliminated, by attentional load during exposure. In Experiment 2, participants saw an ambiguous visual speech gesture that was disambiguated auditorily as either /p/ or /t/. At test, listeners categorized the steps of a visual /p/-/t/ continuum more often in line with the prior exposure. Imposing load in the auditory modality during exposure did not reduce the aftereffect of this type of cross-modal phonetic recalibration. Together, these results suggest that auditory attentional resources are needed for the processing of auditory speech and/or for the shifting of auditory phonetic category boundaries. Listeners thus need to dedicate attentional resources in order to accommodate talker idiosyncrasies in audiovisual speech.

Tolerance for audiovisual asynchrony is enhanced by the spectrotemporal fidelity of the speaker's mouth movements and speech

We examined the relationship between tolerance for audiovisual onset asynchrony (AVOA) and the spectrotemporal fidelity of the spoken words and the speaker's mouth movements. In two experiments that only varied in the temporal order of sensory modality, visual speech leading (exp1) or lagging (exp2) acoustic speech, participants watched intact and blurred videos of a speaker uttering trisyllabic words and nonwords that were noise vocoded with 4-, 8-, 16-, and 32-channels. They judged whether the speaker's mouth movements and the speech sounds were in-sync or out-of-sync. Individuals perceived synchrony (tolerated AVOA) on more trials when the acoustic speech was more speech-like (8 channels and higher vs. 4 channels), and when visual speech was intact than blurred (exp1 only). These findings suggest that enhanced spectrotemporal fidelity of the audiovisual (AV) signal prompts the brain to widen the window of integration promoting the fusion of temporally distant AV percepts.