Towards a cumulative science of vocal markers of autism: a cross-linguistic meta-analysis-based investigation of acoustic markers in American and Danish autistic children

Fusaroli, Riccardo; Grossman, Ruth B.; Bilenberg, Niels; Cantio, Cathriona; Jepsen, Jens Richardt Moellegaard; Weed, Ethan

doi:10.1101/2021.07.13.452165

Cited by 6 publications

(10 citation statements)

References 72 publications

(101 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Even though systematic reviews and meta-analyses provide the framework for combining results of several studies, the obtained results should be taken with caution due to several issues including heterogeneity between studies and publication bias 87 . Furthermore, for the topic of the current study, the spoken language of participants and the task that was used for voice elicitation were different extensively between studies.…”

Section: Discussionmentioning

confidence: 99%

“…Anyway, the outcomes from a systematic review/meta-analysis study can be considered as a starting point in future studies for investigating the effect of potential confounding factors. In this perspective, Fusaroli et al performed a cumulative yet self-correcting approach according to the outcomes of their previous meta-analysis 35 in order to propose guidelines for overcoming the naïve shortcoming of a systematic review/meta-analysis study 87 .…”

Section: Discussionmentioning

confidence: 99%

See 1 more Smart Citation

Distinctive prosodic features of people with autism spectrum disorder: a systematic review and meta-analysis study

Asghari

Farashi

Bashirian

et al. 2021

Sci Rep

View full text Add to dashboard Cite

In this systematic review, we analyzed and evaluated the findings of studies on prosodic features of vocal productions of people with autism spectrum disorder (ASD) in order to recognize the statistically significant, most confirmed and reliable prosodic differences distinguishing people with ASD from typically developing individuals. Using suitable keywords, three major databases including Web of Science, PubMed and Scopus, were searched. The results for prosodic features such as mean pitch, pitch range and variability, speech rate, intensity and voice duration were extracted from eligible studies. The pooled standard mean difference between ASD and control groups was extracted or calculated. Using I2 statistic and Cochrane Q-test, between-study heterogeneity was evaluated. Furthermore, publication bias was assessed using funnel plot and its significance was evaluated using Egger’s and Begg’s tests. Thirty-nine eligible studies were retrieved (including 910 and 850 participants for ASD and control groups, respectively). This systematic review and meta-analysis showed that ASD group members had a significantly larger mean pitch (SMD = − 0.4, 95% CI [− 0.70, − 0.10]), larger pitch range (SMD = − 0.78, 95% CI [− 1.34, − 0.21]), longer voice duration (SMD = − 0.43, 95% CI [− 0.72, − 0.15]), and larger pitch variability (SMD = − 0.46, 95% CI [− 0.84, − 0.08]), compared with typically developing control group. However, no significant differences in pitch standard deviation, voice intensity and speech rate were found between groups. Chronological age of participants and voice elicitation tasks were two sources of between-study heterogeneity. Furthermore, no publication bias was observed during analyses (p > 0.05). Mean pitch, pitch range, pitch variability and voice duration were recognized as the prosodic features reliably distinguishing people with ASD from TD individuals.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Discussionmentioning

confidence: 99%

Distinctive prosodic features of people with autism spectrum disorder: a systematic review and meta-analysis study

Asghari

Farashi

Bashirian

et al. 2021

Sci Rep

View full text Add to dashboard Cite

show abstract

“…Summary features neglecting the temporal dynamics of vocal productions are likely to neglect relevant aspects of autistic voices. While hand-engineering of temporal acoustic features is certainly an option (Fusaroli et al, 2015), recent developments in time-sensitive algorithms – such as recurrent neural networks dealing with the sequence of acoustic values in time - might also be employed. More radically, in the presence of larger datasets, deep learning algorithms have also been successfully used to automatically infer relevant features from the data.…”

Section: Discussionmentioning

confidence: 99%

“…Nevertheless, to mitigate the issue we relied on summary features (median and interquartile range) across each recording and on repeated measures for each participant (2-10 recordings per participant per task). Note that by using summary features of this kind, we are neglecting temporal variation, which are likely to be relevant as markers of autism (Bone et al, 2014; Fusaroli et al, 2015). However, our goal is to assess the generalizability of current ML practices in the field and we leave the identification of better features than those currently used to future work.…”

Section: Methodsmentioning

confidence: 99%

Vocal markers of autism: assessing the generalizability of machine learning models

Rybner

Jessen

Mortensen

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

Machine learning (ML) approaches show increasing promise in their ability to identify vocal markers of autism. Nonetheless, it is unclear to what extent such markers generalize to new speech samples collected e.g., using a different speech task or in a different language. In this paper, we systematically assess the generalizability of ML findings across a variety of contexts. We train promising published ML models of vocal markers of autism on novel cross-linguistic datasets following a rigorous pipeline to minimize overfitting, including cross-validated training and ensemble models. We test the generalizability of the models by testing them on i) different participants from the same study, performing the same task; ii) the same participants, performing a different (but similar) task; iii) a different study with participants speaking a different language, performing the same type of task.While model performance is similar to previously published findings when trained and tested on data from the same study (out-of-sample performance), there is considerable variance between studies. Crucially, the models do not generalize well to different, though similar, tasks and not at all to new languages. The ML pipeline is openly shared.Generalizability of ML models of vocal markers - and more generally biobehavioral markers - of autism is an issue. We outline three recommendations for strategies researchers could take to be more explicit about generalizability and improve it in future studies.

show abstract

“…A recent systematic review and meta-analysis (Fusaroli, Lambrechts, Bang, Bowler, & Gaigg, 2017) highlights several potential acoustic contributors to group differences. The meta-analytic findings point to higher pitch, and increased pitch variability, number of pauses and pause duration, although replication attempts show these patterns might be not be generalizable across languages and samples (see, Fusaroli et al 2021). While some studies have found that acoustic differences map onto expert clinician ratings of autism-specific symptoms (McCann et al, 2007; Study 1, Diehl et al 2009), these findings have not always replicated, even within the same lab (Study 2, ibid;Fusaroli et al 2021).…”

mentioning

confidence: 99%

Different in Different Ways: A Network-Analysis Approach to Voice and Prosody in Autism Spectrum Disorder

Weed

Fusaroli

Simmons

et al. 2023

Language Learning and Development

View full text Add to dashboard Cite

Speakers with autism spectrum disorder (ASD) are commonly reported to have atypical voice and prosody that impact impression formation. While human raters are highly accurate in distinguishing between ASD and neurotypical (NT) speakers, there is little consensus on which specific acoustic features differentiate these groups, suggesting the presence of multiple prosodic profiles. To investigate this possibility, we modelled the speech from a selection of speakers (N = 30), with and without ASD, as a network of nodes defined by acoustic features, and used a community-detection algorithm to identify clusters of speakers who were acoustically similar.Analyses suggested three clusters: one primarily composed of speakers with ASD, one of mostly NT speakers, and one comprised of an even mixture of ASD and NT speakers. Human raters are highly reliable at distinguishing speakers with and without ASD based on perceptual voice and prosodic cues. Our results suggest that community-detection methods using a network approach may complement commonly employed human ratings to improve our understanding of the intonation profiles in ASD.Difficulties with speech and language are a defining feature of autism spectrum disorder (ASD) (Association, 2013), and poor communication skills exacerbate the risk being bullied five-fold (Cappadocia, Weiss, & Pepler, 2012). Atypical prosody and vocal presentation impact social integration and employment opportunities. Indeed, some older research suggests that unusual prosody is among the first features that elicit an impression of oddness from others (Mesibov, 1992;Shriberg & Widder, 1990;Van Bourgondien & Woods, 1992). Even individuals with ASD whose expressive and receptive language scores are in the typical range have difficulty with prosody (Shriberg, Paul, Black, & Santen, 2011), and children with ASD are rated as more socially awkward on the basis of audio speech samples alone (Grossman, 2015;Redford, Kapatsinski, & Cornell-Fabiano, 2018;Sasson et al., 2017). One mechanism suggested for this is that speakers with ASD are more consistent in their phonetic production than NT speakers, as reported in a study of adults with and without ASD (Kissine, Geelhand, Philippart, Harmegnies & Deliens, 2021). Atypical vocal quality is one of the signals of the "frank" autism presentation, apparent to expert clinicians within just a few seconds (Marchena & Miller, 2017). However, despite their significant clinical impact, the specific acoustic features underlying speech differences in ASD are as yet poorly understood. This is a clinically important gap in knowledge, as a better understanding of what acoustic cues contribute to atypical speech quality would enable us to develop better-informed targets for intervention. The goal of this manuscript is to combine information from acoustic analyses with ratings from naïve and clinical raters, in order to provide some insights into what acoustic qualities make a voice distinctively autistic.Although clinicians since Asperger have described the voices of people...

show abstract

Towards a cumulative science of vocal markers of autism: a cross-linguistic meta-analysis-based investigation of acoustic markers in American and Danish autistic children

Cited by 6 publications

References 72 publications

Distinctive prosodic features of people with autism spectrum disorder: a systematic review and meta-analysis study

Distinctive prosodic features of people with autism spectrum disorder: a systematic review and meta-analysis study

Vocal markers of autism: assessing the generalizability of machine learning models

Different in Different Ways: A Network-Analysis Approach to Voice and Prosody in Autism Spectrum Disorder

Contact Info

Product

Resources

About