Comparison of performance of automatic recognizers for stutters in speech trained with event or interval markers

Introduction

Automatic recognition of stutters (ARS) from speech recordings can facilitate objective assessment and intervention for people who stutter. However, the performance of ARS systems may depend on how the speech data are segmented and labelled for training and testing. This study compared two segmentation methods: event-based, which delimits speech segments by their fluency status, and interval-based, which uses fixed-length segments regardless of fluency.

Methods

Machine learning models were trained and evaluated on interval-based and event-based stuttered speech corpora. The models used acoustic and linguistic features extracted from the speech signal and the transcriptions generated by a state-of-the-art automatic speech recognition system.

Results

The results showed that event-based segmentation led to better ARS performance than interval-based segmentation, as measured by the area under the curve (AUC) of the receiver operating characteristic. The results suggest differences in the quality and quantity of the data because of segmentation method. The inclusion of linguistic features improved the detection of whole-word repetitions, but not other types of stutters.

Discussion

The findings suggest that event-based segmentation is more suitable for ARS than interval-based segmentation, as it preserves the exact boundaries and types of stutters. The linguistic features provide useful information for separating supra-lexical disfluencies from fluent speech but may not capture the acoustic characteristics of stutters. Future work should explore more robust and diverse features, as well as larger and more representative datasets, for developing effective ARS systems.

Using Hybrid HMM/DNN Embedding Extractor Models in Computational Paralinguistic Tasks

The field of computational paralinguistics emerged from automatic speech processing, and it covers a wide range of tasks involving different phenomena present in human speech. It focuses on the non-verbal content of human speech, including tasks such as spoken emotion recognition, conflict intensity estimation and sleepiness detection from speech, showing straightforward application possibilities for remote monitoring with acoustic sensors. The two main technical issues present in computational paralinguistics are (1) handling varying-length utterances with traditional classifiers and (2) training models on relatively small corpora. In this study, we present a method that combines automatic speech recognition and paralinguistic approaches, which is able to handle both of these technical issues. That is, we trained a HMM/DNN hybrid acoustic model on a general ASR corpus, which was then used as a source of embeddings employed as features for several paralinguistic tasks. To convert the local embeddings into utterance-level features, we experimented with five different aggregation methods, namely mean, standard deviation, skewness, kurtosis and the ratio of non-zero activations. Our results show that the proposed feature extraction technique consistently outperforms the widely used x-vector method used as the baseline, independently of the actual paralinguistic task investigated. Furthermore, the aggregation techniques could be combined effectively as well, leading to further improvements depending on the task and the layer of the neural network serving as the source of the local embeddings. Overall, based on our experimental results, the proposed method can be considered as a competitive and resource-efficient approach for a wide range of computational paralinguistic tasks.

INFANT CRYING DETECTION IN REAL-WORLD ENVIRONMENTS

Most existing cry detection models have been tested with data collected in controlled settings. Thus, the extent to which they generalize to noisy and lived environments is unclear. In this paper, we evaluate several established machine learning approaches including a model leveraging both deep spectrum and acoustic features. This model was able to recognize crying events with F1 score 0.613 (Precision: 0.672, Recall: 0.552), showing improved external validity over existing methods at cry detection in everyday real-world settings. As part of our evaluation, we collect and annotate a novel dataset of infant crying compiled from over 780 hours of labeled real-world audio data, captured via recorders worn by infants in their homes, which we make publicly available. Our findings confirm that a cry detection model trained on in-lab data underperforms when presented with real-world data (in-lab test F1: 0.656, real-world test F1: 0.236), highlighting the value of our new dataset and model.

Multimodal Emotion Recognition on RAVDESS Dataset Using Transfer Learning

Emotion Recognition is attracting the attention of the research community due to the multiple areas where it can be applied, such as in healthcare or in road safety systems. In this paper, we propose a multimodal emotion recognition system that relies on speech and facial information. For the speech-based modality, we evaluated several transfer-learning techniques, more specifically, embedding extraction and Fine-Tuning. The best accuracy results were achieved when we fine-tuned the CNN-14 of the PANNs framework, confirming that the training was more robust when it did not start from scratch and the tasks were similar. Regarding the facial emotion recognizers, we propose a framework that consists of a pre-trained Spatial Transformer Network on saliency maps and facial images followed by a bi-LSTM with an attention mechanism. The error analysis reported that the frame-based systems could present some problems when they were used directly to solve a video-based task despite the domain adaptation, which opens a new line of research to discover new ways to correct this mismatch and take advantage of the embedded knowledge of these pre-trained models. Finally, from the combination of these two modalities with a late fusion strategy, we achieved 80.08% accuracy on the RAVDESS dataset on a subject-wise 5-CV evaluation, classifying eight emotions. The results revealed that these modalities carry relevant information to detect users’ emotional state and their combination enables improvement of system performance.

COVID-19 and Computer Audition: An Overview on What Speech & Sound Analysis Could Contribute in the SARS-CoV-2 Corona Crisis

At the time of writing this article, the world population is suffering from more than 2 million registered COVID-19 disease epidemic-induced deaths since the outbreak of the corona virus, which is now officially known as SARS-CoV-2. However, tremendous efforts have been made worldwide to counter-steer and control the epidemic by now labelled as pandemic. In this contribution, we provide an overview on the potential for computer audition (CA), i.e., the usage of speech and sound analysis by artificial intelligence to help in this scenario. We first survey which types of related or contextually significant phenomena can be automatically assessed from speech or sound. These include the automatic recognition and monitoring of COVID-19 directly or its symptoms such as breathing, dry, and wet coughing or sneezing sounds, speech under cold, eating behaviour, sleepiness, or pain to name but a few. Then, we consider potential use-cases for exploitation. These include risk assessment and diagnosis based on symptom histograms and their development over time, as well as monitoring of spread, social distancing and its effects, treatment and recovery, and patient well-being. We quickly guide further through challenges that need to be faced for real-life usage and limitations also in comparison with non-audio solutions. We come to the conclusion that CA appears ready for implementation of (pre-)diagnosis and monitoring tools, and more generally provides rich and significant, yet so far untapped potential in the fight against COVID-19 spread.

Acoustic and Language Based Deep Learning Approaches for Alzheimer's Dementia Detection From Spontaneous Speech

Current methods for early diagnosis of Alzheimer's Dementia include structured questionnaires, structured interviews, and various cognitive tests. Language difficulties are a major problem in dementia as linguistic skills break down. Current methods do not provide robust tools to capture the true nature of language deficits in spontaneous speech. Early detection of Alzheimer's Dementia (AD) from spontaneous speech overcomes the limitations of earlier approaches as it is less time consuming, can be done at home, and is relatively inexpensive. In this work, we re-implement the existing NLP methods, which used CNN-LSTM architectures and targeted features from conversational transcripts. Our work sheds light on why the accuracy of these models drops to 72.92% on the ADReSS dataset, whereas, they gave state of the art results on the DementiaBank dataset. Further, we build upon these language input-based recurrent neural networks by devising an end-to-end deep learning-based solution that performs a binary classification of Alzheimer's Dementia from the spontaneous speech of the patients. We utilize the ADReSS dataset for all our implementations and explore the deep learning-based methods of combining acoustic features into a common vector using recurrent units. Our approach of combining acoustic features using the Speech-GRU improves the accuracy by 2% in comparison to acoustic baselines. When further enriched by targeted features, the Speech-GRU performs better than acoustic baselines by 6.25%. We propose a bi-modal approach for AD classification and discuss the merits and opportunities of our approach.

Using Complexity-Identical Human- and Machine-Directed Utterances to Investigate Addressee Detection for Spoken Dialogue Systems

Human-machine addressee detection (H-M AD) is a modern paralinguistics and dialogue challenge that arises in multiparty conversations between several people and a spoken dialogue system (SDS) since the users may also talk to each other and even to themselves while interacting with the system. The SDS is supposed to determine whether it is being addressed or not. All existing studies on acoustic H-M AD were conducted on corpora designed in such a way that a human addressee and a machine played different dialogue roles. This peculiarity influences speakers’ behaviour and increases vocal differences between human- and machine-directed utterances. In the present study, we consider the Restaurant Booking Corpus (RBC) that consists of complexity-identical human- and machine-directed phone calls and allows us to eliminate most of the factors influencing speakers’ behaviour implicitly. The only remaining factor is the speakers’ explicit awareness of their interlocutor (technical system or human being). Although complexity-identical H-M AD is essentially more challenging than the classical one, we managed to achieve significant improvements using data augmentation (unweighted average recall (UAR) = 0.628) over native listeners (UAR = 0.596) and a baseline classifier presented by the RBC developers (UAR = 0.539).

Synchronization in Interpersonal Speech

During both positive and negative dyadic exchanges, individuals will often unconsciously imitate their partner. A substantial amount of research has been made on this phenomenon, and such studies have shown that synchronization between communication partners can improve interpersonal relationships. Automatic computational approaches for recognizing synchrony are still in their infancy. In this study, we extend on previous work in which we applied a novel method utilizing hand-crafted low-level acoustic descriptors and autoencoders (AEs) to analyse synchrony in the speech domain. For this purpose, a database consisting of 394 in-the-wild speakers from six different cultures, is used. For each speaker in the dyadic exchange, two AEs are implemented. Post the training phase, the acoustic features for one of the speakers is tested using the AE trained on their dyadic partner. In this same way, we also explore the benefits that deep representations from audio may have, implementing the state-of-the-art Deep Spectrum toolkit. For all speakers at varied time-points during their interaction, the calculation of reconstruction error from the AE trained on their respective dyadic partner is made. The results obtained from this acoustic analysis are then compared with the linguistic experiments based on word counts and word embeddings generated by our word2vec approach. The results demonstrate that there is a degree of synchrony during all interactions. We also find that, this degree varies across the 6 cultures found in the investigated database. These findings are further substantiated through the use of 4,096 dimensional Deep Spectrum features.

Predicting Fatigue and Psychophysiological Test Performance from Speech for Safety-Critical Environments

Automatic systems for estimating operator fatigue have application in safety-critical environments. A system which could estimate level of fatigue from speech would have application in domains where operators engage in regular verbal communication as part of their duties. Previous studies on the prediction of fatigue from speech have been limited because of their reliance on subjective ratings and because they lack comparison to other methods for assessing fatigue. In this paper, we present an analysis of voice recordings and psychophysiological test scores collected from seven aerospace personnel during a training task in which they remained awake for 60 h. We show that voice features and test scores are affected by both the total time spent awake and the time position within each subject’s circadian cycle. However, we show that time spent awake and time-of-day information are poor predictors of the test results, while voice features can give good predictions of the psychophysiological test scores and sleep latency. Mean absolute errors of prediction are possible within about 17.5% for sleep latency and 5–12% for test scores. We discuss the implications for the use of voice as a means to monitor the effects of fatigue on cognitive performance in practical applications.