Speech Task Research Articles

Importance of Task Selection for Connected Speech Analysis in Patients with Alzheimer’s Disease from an Ethnically Diverse Sample

Features of linguistic impairment in Alzheimer’s disease (AD) are primarily derived from English-speaking patients. Little is known regarding such deficits in linguistically diverse speakers with AD. We aimed to detail linguistic profiles (speech rate, dysfluencies, syntactic, lexical, morphological, semantics) from two connected speech tasks–Frog Story and picture description–in Bengali-speaking AD patients. The Frog Story detected group differences on all six linguistic levels, compared to only three with picture description. Critically, Frog Story captured the language-specific differences between the groups. Careful consideration should be given to the choice of connected speech tasks for dementia diagnosis in linguistically diverse populations.

Socioeconomic status in early adolescence predicts blunted stress responses in adulthood.

Individuals who grow up in families with lower socioeconomic status (SES) tend to experience disproportionate rates of chronic stress. The “freeze” response, characterized by blunted cardiovascular reactivity and reduced engagement with the environment, is associated with chronic stress and may be utilized when an individual is unable to escape or overcome environmental stressors. Using a diverse community sample of 184 adolescents followed from the age of 13 to 29 years, along with their friends and romantic partners, this study examined links between family SES and stress responses in adulthood. Low family SES at the age of 13 years directly predicted blunted heart rate responding and fewer attempts to answer math problems during a modified version of the Trier Social Stress Task at the age of 29 years. Indirect effects were found from low family SES to blunted respiratory sinus arrhythmia responding and the number of words spoken during a speech task. SES at the age of 29 years mediated many of these relations. Findings held after accounting for a number of potential confounds, including adolescent academic and attachment functioning and body mass index. We interpret these findings as evidence that low familial SES may predict freezing‐type responses in adulthood.

Detection and differentiation of ataxic and hypokinetic dysarthria in cerebellar ataxia and parkinsonian disorders via wave splitting and integrating neural networks.

Dysarthria may present during the natural course of many degenerative neurological conditions. Hypokinetic and ataxic dysarthria are common in movement disorders and represent the underlying neuropathology. We developed an artificial intelligence (AI) model to distinguish ataxic dysarthria and hypokinetic dysarthria from normal speech and differentiate ataxic and hypokinetic speech in parkinsonian diseases and cerebellar ataxia. We screened 804 perceptual speech analyses performed in the Samsung Medical Center Neurology Department between January 2017 and December 2020. The data of patients diagnosed with parkinsonian disorders or cerebellar ataxia were included. Two speech tasks (numbering from 1 to 50 and reading nine sentences) were analyzed. We adopted convolutional neural networks and developed a patch-wise wave splitting and integrating AI system for audio classification (PWSI-AI-AC) to differentiate between ataxic and hypokinetic speech. Of the 395 speech recordings for the reading task, 76, 112, and 207 were from normal, ataxic dysarthria, and hypokinetic dysarthria subjects, respectively. Of the 409 recordings of the numbering task, 82, 111, and 216 were from normal, ataxic dysarthria, and hypokinetic dysarthria subjects, respectively. The reading and numbering task recordings were classified with 5-fold cross-validation using PWSI-AI-AC as follows: hypokinetic dysarthria vs. others (area under the curve: 0.92 ± 0.01 and 0.92 ± 0.02), ataxia vs. others (0.93 ± 0.04 and 0.89 ± 0.02), hypokinetic dysarthria vs. ataxia (0.96 ± 0.02 and 0.95 ± 0.01), hypokinetic dysarthria vs. none (0.86 ± 0.03 and 0.87 ± 0.05), and ataxia vs. none (0.87 ± 0.07 and 0.87 ± 0.09), respectively. PWSI-AI-AC showed reliable performance in differentiating ataxic and hypokinetic dysarthria and effectively augmented data to classify the types even with limited training samples. The proposed fully automatic AI system outperforms neurology residents. Our model can provide effective guidelines for screening related diseases and differential diagnosis of neurodegenerative diseases.

Protocol for Rhapsody: a longitudinal observational study examining the feasibility of speech phenotyping for remote assessment of neurodegenerative and psychiatric disorders

IntroductionNeurodegenerative and psychiatric disorders (NPDs) confer a huge health burden, which is set to increase as populations age. New, remotely delivered diagnostic assessments that can detect early stage NPDs by...

Clinical Feasibility of Speech Phenotyping for Remote Assessment of Neurodegenerative and Psychiatric Disorders (RHAPSODY): a study protocol

IntroductionThe diagnosis of neurodegenerative and psychiatric disorders (NPDs) in primary care can suffer from inefficiencies resulting in misdiagnoses and delayed diagnosis, limiting effective treatment options. The development of speech and language-based profiling biomarkers could aid in achieving earlier motor diagnosis for PD for instance, or more accurate diagnosis of clinically similar or late presenting NPDs.ObjectivesRHAPSODY aims to investigate the feasibility of the remote administration of a battery of speech tasks in eliciting continuous narrative speech across a range of NPDs. The project also aims to determine the feasibility of using acoustic and linguistic biomarkers from speech data to support the clinical assessment and disambiguation of common NPDsMethodsAll participants (n=250) will take part in a single virtual telemedicine video conference with a researcher in which they are screened and complete a battery of speech tasks, in addition to cohort-specific screening measures. Over the following month, participants will be asked to complete a series of short, self-administered speech assessments via a smartphone application.ResultsThe speech tasks will be audio-recorded and analysed on Novoic’s technology platform. Objectives will be analysed using measures including average length of speech elicitation for speech tasks, intra- and inter-subject variance, differences in linguistic patterns, and response rates to speech assessments.ConclusionsThe analyses could help to identify and validate speech-derived clinical biomarkers to support clinicians in detecting and disambiguating between NPDs with heterogeneous presentations. This should further support earlier intervention, improved treatment options and improved quality of life.DisclosureIn terms of significant financial interest and relationships, it is emphasised that the private organisation Novoic, who aim to develop speech algorithms for diagnostic use, is the study’s sponsor and employees or former employees of this company comprise

Remote Assessment of Disease and Relapse in Major Depressive Disorder (RADAR-MDD): Recruitment, retention, and data availability in a longitudinal remote measurement study

IntroductionMajor Depressive Disorder (MDD) is prevalent, often chronic, and requires ongoing monitoring of symptoms to track response to treatment and identify early indicators of relapse. Remote Measurement Technologies (RMT) provide an exciting opportunity to transform the measurement and management of MDD, via data collected from inbuilt smartphone sensors and wearable devices alongside app-based questionnaires and tasks.ObjectivesTo describe the amount of data collected during a multimodal longitudinal RMT study, in an MDD population.MethodsRADAR-MDD is a multi-centre, prospective observational cohort study. People with a history of MDD were provided with a wrist-worn wearable, and several apps designed to: a) collect data from smartphone sensors; and b) deliver questionnaires, speech tasks and cognitive assessments and followed-up for a maximum of 2 years.ResultsA total of 623 individuals with a history of MDD were enrolled in the study with 80% completion rates for primary outcome assessments across all timepoints. 79.8% of people participated for the maximum amount of time available and 20.2% withdrew prematurely. Data availability across all RMT data types varied depending on the source of data and the participant-burden for each data type. We found no evidence of an association between the severity of depression symptoms at baseline and the availability of data. 110 participants had > 50% data available across all data types, and thus able to contribute to multiparametric analyses.ConclusionsRADAR-MDD is the largest multimodal RMT study in the field of mental health. Here, we have shown that collecting RMT data from a clinical population is feasible.DisclosureNo significant relationships.

Neuromorphic Time-Multiplexed Reservoir Computing With On-the-Fly Weight Generation for Edge Devices.

The human brain has evolved to perform complex and computationally expensive cognitive tasks, such as audio-visual perception and object detection, with ease. For instance, the brain can recognize speech in different dialects and perform other cognitive tasks, such as attention, memory, and motor control, with just 20 W of power consumption. Taking inspiration from neural systems, we propose a low-power neuromorphic hardware architecture to perform classification on temporal data at the edge. The proposed architecture uses a neuromorphic cochlea model for feature extraction and reservoir computing (RC) framework as a classifier. In the proposed hardware architecture, the RC framework is modified for on-the-fly generation of reservoir connectivity, along with binary feedforward and reservoir weights. Also, a large reservoir is split into multiple small reservoirs for efficient use of hardware resources. These modifications reduce the computational and memory resources required, thereby resulting in a lower power budget. The proposed classifier is validated for speech and human activity recognition (HAR) tasks. We have prototyped our hardware architecture using Intel's cyclone-10 low-power series field-programmable gate array (FPGA), consuming only 4790 logic elements (LEs) and 34.9-kB memory, making it a perfect candidate for edge computing applications. Moreover, we have implemented a complete system for speech recognition with the feature extraction block (cochlea model) and the proposed classifier, utilizing 15 532 LEs and 38.4-kB memory. By using the proposed idea of multiple small reservoirs along with on-the-fly generation of reservoir binary weights, our architecture can reduce the power consumption and memory requirement by order of magnitude compared to existing FPGA models for speech recognition tasks with similar complexity.

Measuring neuropsychiatric symptoms in early dementia patients using speech analysis

IntroductionCertain neuropsychiatric symptoms (NPS), namely apathy, depression and anxiety demonstrated great value in predicting dementia progression representing eventually an opportunity window for timely diagnosis and treatment. However, sensitive and objective markers of these symptoms are still missing.ObjectivesTo investigate the association between automatically extracted speech features and NPS in early-stage dementia patients.MethodsSpeech of 141 patients aged 65 or older with neurocognitive disorder was recorded while performing two short narrative speech tasks. Presence of NPS was assessed by the Neuropsychiatric Inventory. Paralinguistic markers relating to prosodic, formant, source, and temporal qualities of speech were automatically extracted, correlated with NPS. Machine learning experiments were carried out to validate the diagnostic power of extracted markers.ResultsDifferent speech variables seem to be associated with specific neuropsychiatric symptoms of dementia; apathy correlates with temporal aspects, anxiety with voice quality and this was mostly consistent between male and female after correction for cognitive impairment. Machine learning regressors are able to extract information from speech features and perform above baseline in predicting anxiety, apathy and depression scores.ConclusionsDifferent NPS seem to be characterized by distinct speech features which in turn were easily extractable automatically from short vocal tasks. These findings support the use of speech analysis for detecting subtypes of NPS. This could have great implications for future clinical trials.DisclosureNo significant relationships.

Motor learning principles in the service of speech disorders

IntroductionChildhood Apraxia of Speech (CAS) was declared as a motor speech disorder by ASHA (2007). Yet, until then it was mainly addressed as a phonological disorder and until these days, 14 years later, the treatment of CAS is yet to be motor based worldwide. Professionals finds it hard to diagnose it clearly due to comorbidity with communication and language disorders.ObjectivesThis non clarity might lead to non-accurate treatment since the essence of the syndrome is not addressed. An accurate treatment will integrate knowledge from several domains: communication, Language, Sensory, behavioural, emotional, cognitive and, the most important one for CAS, motor learning.MethodsMotor learning is an area of knowledge which is learnt usually in sport academy, while Its main practical purpose is to improve training methods in sport. The use of motor learning knowledge doesn’t belong to the world of sport primarily but rather to the world of movement wherever it exists. One of the fascinating areas of movement is speech.Results Speech in its basic form is motor based, before it being used as a motor tool for language and communication. It is the most complicated motor task in the human body since for every syllable we activate directly and indirectly over 100 muscles. The children who can’t acquire speech spontaneously due to severe deficit in motor planning, need to practice motor speech tasks repeatedly and accurately.ConclusionsThis lecture will present the use of 20 motor learning principles in the speech treatment via the VML methodDisclosureI am the founder of the VML method while teaching it in various countries

State anxiety and frontal alpha asymmetry effects of the REThink online video game for children and adolescents: A six-month follow-up

BackgroundEmotional disorders are the most prevalent mental health conditions affecting children and adolescents. Thus, it becomes essential to develop and test early intervention strategies that are accessible and attractive as therapeutic strategies and can effectively improve youth's emotional and psychobiological reactivity to distress. MethodsA randomized control trial compared the prevention effects of a newly-developed therapeutic game based on Rational Emotive Behavior Education, REThink, to those of a standard, face-to-face group Rational Emotive Behavior Education (REBE) and those of a waitlist. Out of 142 healthy children and adolescents who completed the intervention stage, 137 completed follow-up assessment measuring subjective anxiety and biological stress reactivity in an impromptu speech task. ResultsNo differences were found between groups at follow-up in terms of subjective state anxiety variations between baseline, anticipation, speech and recovery. In terms of the psychobiological index, results showed an increase in left asymmetry for the REThink group at post-test and follow-up compared to pre-intervention levels. The same trend was found for the WL group, but not for the REBE group. ConclusionsPreliminary results suggest that decreases in the biological reactivity can be maintained in the long run following a therapeutic video game for children and adolescents. Future research needs to clarify the adequate usage of the therapeutic videogames to bring significant improvements in the psychobiological functioning of the youths.

SPEECH TASKS RELEVANT TO SLEEPINESS DETERMINED WITH DEEP TRANSFER LEARNING.

Excessive sleepiness in attention-critical contexts can lead to adverse events, such as car crashes. Detecting and monitoring sleepiness can help prevent these adverse events from happening. In this paper, we use the Voiceome dataset to extract speech from 1,828 participants to develop a deep transfer learning model using Hidden-Unit BERT (HuBERT) speech representations to detect sleepiness from individuals. Speech is an under-utilized source of data in sleep detection, but as speech collection is easy, cost-effective, and non-invasive, it provides a promising resource for sleepiness detection. Two complementary techniques were conducted in order to seek converging evidence regarding the importance of individual speech tasks. Our first technique, masking, evaluated task importance by combining all speech tasks, masking selected responses in the speech, and observing systematic changes in model accuracy. Our second technique, separate training, compared the accuracy of multiple models, each of which used the same architecture, but was trained on a different subset of speech tasks. Our evaluation shows that the best-performing model utilizes the memory recall task and categorical naming task from the Boston Naming Test, which achieved an accuracy of 80.07% (F1-score of 0.85) and 81.13% (F1-score of 0.89), respectively.

A Practice-oriented Intervention Study on the Use of Conjunctional Adverbs

Conjunctional adverbs and formulaic expressions are the fundamental mechanisms in the development of fluent speech, providing coherence and cohesion within discourse, but also contributing to perceived fluency or the listener’s impression of native-like speech. Research has emphasized the exceptional importance of explicitly teaching those lexical units that directly affect the development of speech fluency since they are very important in all phases during the process of speech production. The present study aims to expand the existing knowledge by providing insight into how formal teaching instruction can contribute to the automatization of different conjunctional adverbs to achieve coherence and contribute to perceived fluency. A group of eight students of German studies at the Faculty of Philosophy of Humanities and Social Sciences in Split recorded a spontaneous dialogue based on three given topics twice, before and after the intervention. The intervention included different speech tasks with the implementation of a variety of native like conjunctional adverbs (e.g., außerdem, darüberhinaus, zudem, dazu, zusätzlich, daher, nämlich, deswegen, etc.). The results of the study have mostly shown that learners tend to employ an unsatisfactory repertoire of a variety of lexical units with a narrower range of different functions. In other words, an extreme overuse of the discourse markers und and also was recorded in the first dialogue. The analysis of the second recording after the intervention displayed no changes.

The Effect of Right Temporal Lobe Gliomas on Left and Right Hemisphere Neural Processing During Speech Perception and Production Tasks.

Using fMRI, we investigated how right temporal lobe gliomas affecting the posterior superior temporal sulcus alter neural processing observed during speech perception and production tasks. Behavioural language testing showed that three pre-operative neurosurgical patients with grade 2, grade 3 or grade 4 tumours had the same pattern of mild language impairment in the domains of object naming and written word comprehension. When matching heard words for semantic relatedness (a speech perception task), these patients showed under-activation in the tumour infiltrated right superior temporal lobe compared to 61 neurotypical participants and 16 patients with tumours that preserved the right postero-superior temporal lobe, with enhanced activation within the (tumour-free) contralateral left superior temporal lobe. In contrast, when correctly naming objects (a speech production task), the patients with right postero-superior temporal lobe tumours showed higher activation than both control groups in the same right postero-superior temporal lobe region that was under-activated during auditory semantic matching. The task dependent pattern of under-activation during the auditory speech task and over-activation during object naming was also observed in eight stroke patients with right hemisphere infarcts that affected the right postero-superior temporal lobe compared to eight stroke patients with right hemisphere infarcts that spared it. These task-specific and site-specific cross-pathology effects highlight the importance of the right temporal lobe for language processing and motivate further study of how right temporal lobe tumours affect language performance and neural reorganisation. These findings may have important implications for surgical management of these patients, as knowledge of the regions showing functional reorganisation may help to avoid their inadvertent damage during neurosurgery.

Using digital health tools for the Remote Assessment of Treatment Prognosis in Depression (RAPID): a study protocol for a feasibility study

IntroductionDigital health tools such as smartphones and wearable devices could improve psychological treatment outcomes in depression through more accurate and comprehensive measures of patient behaviour. However, in this emerging field,...

Spiking Cochlea With System-Level Local Automatic Gain Control

Including local automatic gain control (AGC) circuitry into a silicon cochlea design has been challenging because of transistor mismatch and model complexity. To address this, we present an alternative system-level algorithm that implements channel-specific AGC in a silicon spiking cochlea by measuring the output spike activity of individual channels. The bandpass filter gain of a channel is adapted dynamically to the input amplitude so that the average output spike rate stays within a defined range. Because this AGC mechanism only needs counting and adding operations, it can be implemented at low hardware cost in a future design. We evaluate the impact of the local AGC algorithm on a classification task where the input signal varies over 32dB input range. Two classifier types receiving cochlea spike features were tested on a speech versus noise classification task. The logistic regression classifier achieves an average of 6% improvement and 40.8% relative improvement in accuracy when the AGC is enabled. The deep neural network classifier shows a similar improvement for the AGC case and achieves a higher mean accuracy of 96% compared to the best accuracy of 91% from the logistic regression classifier.

Objective and Subjective Benefit of Direct-to-Consumer Hearing Devices in Middle-Aged Adults.

The purpose of this project was to assess subjective and objective benefit provided by several direct-to-consumer hearing devices for middle-aged adults. The primary goal of this study was to determine the extent to which this type of device can yield benefit when users are listening in a range of acoustic conditions, rather than to compare benefit among devices. Participants (M age = 58 years, n = 40) completed a speech perception task with and without one of four direct-to-consumer devices. Stimuli were presented with three types of maskers (steady-state noise, modulated noise, and competing talkers) at two different signal-to-noise ratios and two target levels. Participants also rated the effort required to complete the task with and without the devices and completed a short questionnaire about device comfort and perceived effectiveness. The amount of objective benefit (in terms of speech recognition accuracy) varied among the four products, with two of the four devices yielding statistically significant benefit with medium-to-large effect sizes. Reduction in self-rated listening effort was noted from the use of all four devices, with a moderate effect size. Degree of hearing loss (4-frequency pure-tone average) was not significantly associated with the amount of either subjective or objective benefit. Responses to the posttask questionnaire indicated that many of the participants would be willing to use these or similar devices in the "real world." Our findings support the concept that direct-to-consumer hearing devices have the potential to improve objective and/or subjective speech recognition in middle-aged adults, at least when fit to prescriptive targets. Benefit from these devices was not related to degree of hearing loss in this sample of participants.

Speech-induced action myoclonus

BackgroundSpeech-induced action myoclonus may occur as a component of a generalized myoclonus syndrome. However, it may also present in isolation, or with a paucity of other findings, and be diagnostically challenging. ObjectivesTo report a retrospective case series of restricted speech-induced action myoclonus. MethodsWe reviewed cases of speech-induced action myoclonus evaluated at Mayo Clinic Rochester from 1989 to 2020. We eliminated cases where a more generalized myoclonic disorder was also present. Clinical, imaging, and electrophysiologic data were extracted. ResultsFour cases were identified in which speech-induced action myoclonus of craniofacial muscles was the predominant clinical presentation. All described cranial muscle twitching induced by speaking, and two cases also reported speech interruptions. Diagnosis was confirmed by expert speech pathologists in all cases. Diagnostic aids included modulation with different speech tasks and speaking rates, and surface electrophysiology which confirmed craniofacial myoclonus induced by speaking tasks (three cases). Previous misdiagnosis included functional, dystonic, neuromuscular junction pathology, or hemifacial spasm. Two cases had isolated speech-induced myoclonus, and the other two had coexistent upper limb tremor. Potential etiologic factors were identified in three cases – medication (2), epilepsy (1) – while in one patient no cause was identified. One patient partially improved with anti-myoclonic medication and speech therapy. ConclusionsSpeech-induced action myoclonus may occur in isolation and is frequently misdiagnosed. Diagnostic aids include modulation with different speech tasks and speaking rates, and surface electrophysiology.

Trajectories of socially anxious behavior from age 5 to 13: Temperamental and sociocognitive pathways.

The authors examined temperamental and sociocognitive predictors of socially anxious behavior from preschool to early adolescence. Children (N=227; 59% male; 69% White) completed a speech task at ages 5, 7, 10, and 13 and socially anxious behaviors were coded. Behavioral inhibition (BI) was assessed at ages 2/3 and Theory of Mind (ToM) was assessed at age 4. Data collection occurred between 2003 and 2016. Three trajectories of socially anxious behavior were identified: high stable, average increasing, and low stable. Higher BI was related to the high stable trajectory, whereas lower ToM was related to the increasing trajectory of socially anxious behavior. There are heterogenous pathways of socially anxious behavior, which may be uniquely influenced by early temperamental and sociocognitive factors.

The function of auditory cortex in the elderly using functional near-infrared spectroscopy technology

Objective: Functional near-infrared spectroscopy (fNIRS) was used to study the effect of aging on the neuroimaging characteristics of cerebral cortex in the process of speech perception. Method: Thirty-four adults with normal hearing were recruited from March 2021 to June 2021, including 17 in the young group, with 6 males, 11 females, age (32.1±5.0) years, age range 20-39 years. and 17 in the elderly group, with 6 males, 11 females, age (63.2±2.8) years, age range 60-70 years. The test material was the sentence table of the Mandarin Hearing Test in Noise (MHINT). The task state block experiment design was adopted, and the temporal lobe, Broca's area, Wernicke's area, motor cortex were used as regions of interest. Objective brain imaging technology (fNIRS) combined with subjective psychophysical testing method was used to analyze the activation area and degree of cerebral cortex related to auditory speech perception in the elderly and young people under different listening conditions (quiet, signal-to-noise ratio of 10 dB, 5 dB, 0 dB, -5 dB). SPSS 23 software was used for statistical analysis. Result: The activation area and degree of activation in the elderly group were lower than those in the young group under each task condition; The number of activation channels in the young group were significantly more than those in the old group, and the number of activation channels in the left hemisphere were more than those in the right hemisphere, but there was no difference in the number of activation channels. There were more channels affected by age in the left hemisphere than in the right hemisphere. The activation degree of the young group when the signal-to-noise ratio was 0 dB was significantly higher than that of other signal-to-noise ratio conditions (P<0.05), but there was no significant difference in the old group under the five conditions (P>0.05). The speech recognition score of the young group was higher than that of the old group under all conditions. When the quiet and signal-to-noise ratio was 10 dB, the correct score of the two groups was equal or close to 100%. With the gradual decrease of signal-to-noise ratio, there was a significant difference between the two groups when the signal-to-noise ratio was 5 dB. The speech recognition accuracy of the young group decreased significantly when the signal-to-noise ratio was 0 dB, while that of the old group decreased significantly when the signal-to-noise ratio was 5 dB. Conclusions: With the increase of age, the speech perception in noisy environment and the activity of cerebral cortex gradually deteriorate, and the speech dominance hemisphere (left hemisphere) will be significantly affected by aging. The overall activation area and activation degree of the elderly under different speech tasks are lower than those of the young.

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

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, for example, 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 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. LAY SUMMARY: Machine learning approaches promise to be able to identify autism from voice only. These models underestimate how diverse the contexts in which we speak are, how diverse the languages used are and how diverse autistic voices are. Machine learning approaches need to be more careful in defining their limits and generalizability.