The UK COVID-19 Vocal Audio Dataset is designed for the training and evaluation of machine learning models that classify SARS-CoV-2 infection status or associated respiratory symptoms using vocal audio. The UK Health Security Agency recruited voluntary participants through the national Test and Trace programme and the REACT-1 survey in England from March 2021 to March 2022, during dominant transmission of the Alpha and Delta SARS-CoV-2 variants and some Omicron variant sublineages. Audio recordings of volitional coughs, exhalations, and speech were collected in the \'Speak up and help beat coronavirus\' digital survey alongside demographic, symptom and self-reported respiratory condition data. Digital survey submissions were linked to SARS-CoV-2 test results. The UK COVID-19 Vocal Audio Dataset represents the largest collection of SARS-CoV-2 PCR-referenced audio recordings to date. PCR results were linked to 70,565 of 72,999 participants and 24,105 of 25,706 positive cases. Respiratory symptoms were reported by 45.6% of participants. This dataset has additional potential uses for bioacoustics research, with 11.3% participants self-reporting asthma, and 27.2% with linked influenza PCR test results.

Existing end-to-end speech recognition methods typically employ hybrid decoders based on CTC and Transformer. However, the issue of error accumulation in these hybrid decoders hinders further improvements in accuracy. Additionally, most existing models are built upon Transformer architecture, which tends to be complex and unfriendly to small datasets. Hence, we propose a Nonlinear Regularization Decoding Method for Speech Recognition. Firstly, we introduce the nonlinear Transformer decoder, breaking away from traditional left-to-right or right-to-left decoding orders and enabling associations between any characters, mitigating the limitations of Transformer architectures on small datasets. Secondly, we propose a novel regularization attention module to optimize the attention score matrix, reducing the impact of early errors on later outputs. Finally, we introduce the tiny model to address the challenge of overly large model parameters. The experimental results indicate that our model demonstrates good performance. Compared to the baseline, our model achieves recognition improvements of 0.12%, 0.54%, 0.51%, and 1.2% on the Aishell1, Primewords, Free ST Chinese Corpus, and Common Voice 16.1 datasets of Uyghur, respectively.

Accommodating talker variability is a complex and multi-layered cognitive process. It involves shifting attention to the vocal characteristics of the talker as well as the linguistic content of their speech. Due to an interdependence between voice and phonological processing, multi-talker environments typically incur additional processing costs compared to single-talker environments. A failure or inability to efficiently distribute attention over multiple acoustic cues in the speech signal may have detrimental language learning consequences. Yet, no studies have examined effects of multi-talker processing in populations with atypical perceptual, social and language processing for communication, including autistic people. Employing a classic word-monitoring task, we investigated effects of talker variability in Australian English autistic (n = 24) and non-autistic (n = 28) adults. Listeners responded to target words (e.g., apple, duck, corn) in randomised sequences of words. Half of the sequences were spoken by a single talker and the other half by multiple talkers. Results revealed that autistic participants\' sensitivity scores to accurately-spotted target words did not differ to those of non-autistic participants, regardless of whether they were spoken by a single or multiple talkers. As expected, the non-autistic group showed the well-established processing cost associated with talker variability (e.g., slower response times). Remarkably, autistic listeners\' response times did not differ across single- or multi-talker conditions, indicating they did not show perceptual processing costs when accommodating talker variability. The present findings have implications for theories of autistic perception and speech and language processing.

Lip language recognition urgently needs wearable and easy-to-use interfaces for interference-free and high-fidelity lip-reading acquisition and to develop accompanying data-efficient decoder-modeling methods. Existing solutions suffer from unreliable lip reading, are data hungry, and exhibit poor generalization. Here, we propose a wearable lip language decoding technology that enables interference-free and high-fidelity acquisition of lip movements and data-efficient recognition of fluent lip language based on wearable motion capture and continuous lip speech movement reconstruction. The method allows us to artificially generate any wanted continuous speech datasets from a very limited corpus of word samples from users. By using these artificial datasets to train the decoder, we achieve an average accuracy of 92.0% across individuals (n = 7) for actual continuous and fluent lip speech recognition for 93 English sentences, even observing no training burn on users because all training datasets are artificially generated. Our method greatly minimizes users\' training/learning load and presents a data-efficient and easy-to-use paradigm for lip language recognition.

Hearing one\'s own speech allows for acoustic self-monitoring in real time. Left-hemisphere motor planning regions are thought to give rise to efferent predictions that can be compared to true feedback in sensory cortices, resulting in neural suppression commensurate with the degree of overlap between predicted and actual sensations. Sensory prediction errors thus serve as a possible mechanism of detection of deviant speech sounds, which can then feed back into corrective action, allowing for online control of speech acoustics. The goal of this study was to assess the integrity of this detection-correction circuit in persons with aphasia (PWA) whose left-hemisphere lesions may limit their ability to control variability in speech output. We recorded magnetoencephalography (MEG) while 15 PWA and age-matched controls spoke monosyllabic words and listened to playback of their utterances. From this, we measured speaking-induced suppression of the M100 neural response and related it to lesion profiles and speech behavior. Both speaking-induced suppression and cortical sensitivity to deviance were preserved at the group level in PWA. PWA with more spared tissue in pars opercularis had greater left-hemisphere neural suppression and greater behavioral correction of acoustically deviant pronunciations, whereas sparing of superior temporal gyrus was not related to neural suppression or acoustic behavior. In turn, PWA who made greater corrections had fewer overt speech errors in the MEG task. Thus, the motor planning regions that generate the efferent prediction are integral to performing corrections when that prediction is violated.

Slow cortical oscillations play a crucial role in processing the speech amplitude envelope, which is perceived atypically by children with developmental dyslexia. Here we use electroencephalography (EEG) recorded during natural speech listening to identify neural processing patterns involving slow oscillations that may characterize children with dyslexia. In a story listening paradigm, we find that atypical power dynamics and phase-amplitude coupling between delta and theta oscillations characterize dyslexic versus other child control groups (typically-developing controls, other language disorder controls). We further isolate EEG common spatial patterns (CSP) during speech listening across delta and theta oscillations that identify dyslexic children. A linear classifier using four delta-band CSP variables predicted dyslexia status (0.77 AUC). Crucially, these spatial patterns also identified children with dyslexia when applied to EEG measured during a rhythmic syllable processing task. This transfer effect (i.e., the ability to use neural features derived from a story listening task as input features to a classifier based on a rhythmic syllable task) is consistent with a core developmental deficit in neural processing of speech rhythm. The findings are suggestive of distinct atypical neurocognitive speech encoding mechanisms underlying dyslexia, which could be targeted by novel interventions.

Amyotrophic lateral sclerosis (ALS) is an idiopathic, fatal, and fast-progressive neurodegenerative disease characterized by the degeneration of motor neurons. ALS patients often experience an initial misdiagnosis or a diagnostic delay due to the current unavailability of an efficient biomarker. Since impaired speech is typical in ALS, we hypothesized that functional differences between healthy and ALS participants during speech tasks can be explained by cortical pattern changes, thereby leading to the identification of a neural biomarker for ALS. In this pilot study, we collected magnetoencephalography (MEG) recordings from three early-diagnosed patients with ALS and three healthy controls during imagined (covert) and overt speech tasks. First, we computed sensor correlations, which showed greater correlations for speakers with ALS than healthy controls. Second, we compared the power of the MEG signals in canonical bands between the two groups, which showed greater dissimilarity in the beta band for ALS participants. Third, we assessed differences in functional connectivity, which showed greater beta band connectivity for ALS than healthy controls. Finally, we performed single-trial classification, which resulted in highest performance with beta band features (∼ 98%). These findings were consistent across trials, phrases, and participants for both imagined and overt speech tasks. Our preliminary results indicate that speech-evoked beta oscillations could be a potential neural biomarker for diagnosing ALS. To our knowledge, this is the first demonstration of the detection of ALS from single-trial neural signals.

Human language relies on the correct processing of syntactic information, as it is essential for successful communication between speakers. As an abstract level of language, syntax has often been studied separately from the physical form of the speech signal, thus often masking the interactions that can promote better syntactic processing in the human brain. However, behavioral and neural evidence from adults suggests the idea that prosody and syntax interact, and studies in infants support the notion that prosody assists language learning. Here we analyze a MEG dataset to investigate how acoustic cues, specifically prosody, interact with syntactic representations in the brains of native English speakers. More specifically, to examine whether prosody enhances the cortical encoding of syntactic representations, we decode syntactic phrase boundaries directly from brain activity, and evaluate possible modulations of this decoding by the prosodic boundaries. Our findings demonstrate that the presence of prosodic boundaries improves the neural representation of phrase boundaries, indicating the facilitative role of prosodic cues in processing abstract linguistic features. This work has implications for interactive models of how the brain processes different linguistic features. Future research is needed to establish the neural underpinnings of prosody-syntax interactions in languages with different typological characteristics.

Goal: As an essential human-machine interactive task, emotion recognition has become an emerging area over the decades. Although previous attempts to classify emotions have achieved high performance, several challenges remain open: 1) How to effectively recognize emotions using different modalities remains challenging. 2) Due to the increasing amount of computing power required for deep learning, how to provide real-time detection and improve the robustness of deep neural networks is important. Method: In this paper, we propose a deep learning-based multimodal emotion recognition (MER) called Deep-Emotion, which can adaptively integrate the most discriminating features from facial expressions, speech, and electroencephalogram (EEG) to improve the performance of the MER. Specifically, the proposed Deep-Emotion framework consists of three branches, i.e., the facial branch, speech branch, and EEG branch. Correspondingly, the facial branch uses the improved GhostNet neural network proposed in this paper for feature extraction, which effectively alleviates the overfitting phenomenon in the training process and improves the classification accuracy compared with the original GhostNet network. For work on the speech branch, this paper proposes a lightweight fully convolutional neural network (LFCNN) for the efficient extraction of speech emotion features. Regarding the study of EEG branches, we proposed a tree-like LSTM (tLSTM) model capable of fusing multi-stage features for EEG emotion feature extraction. Finally, we adopted the strategy of decision-level fusion to integrate the recognition results of the above three modes, resulting in more comprehensive and accurate performance. Result and Conclusions: Extensive experiments on the CK+, EMO-DB, and MAHNOB-HCI datasets have demonstrated the advanced nature of the Deep-Emotion method proposed in this paper, as well as the feasibility and superiority of the MER approach.

Language is rooted in our ability to compose: We link words together, fusing their meanings. Links are not limited to neighboring words but often span intervening words. The ability to process these non-adjacent dependencies (NADs) conflicts with the brain\'s sampling of speech: We consume speech in chunks that are limited in time, containing only a limited number of words. It is unknown how we link words together that belong to separate chunks. Here, we report that we cannot-at least not so well. In our electroencephalography (EEG) study, 37 human listeners learned chunks and dependencies from an artificial grammar (AG) composed of syllables. Multi-syllable chunks to be learned were equal-sized, allowing us to employ a frequency-tagging approach. On top of chunks, syllable streams contained NADs that were either confined to a single chunk or crossed a chunk boundary. Frequency analyses of the EEG revealed a spectral peak at the chunk rate, showing that participants learned the chunks. NADs that cross boundaries were associated with smaller electrophysiological responses than within-chunk NADs. This shows that NADs are processed readily when they are confined to the same chunk, but not as well when crossing a chunk boundary. Our findings help to reconcile the classical notion that language is processed incrementally with recent evidence for discrete perceptual sampling of speech. This has implications for language acquisition and processing as well as for the general view of syntax in human language.