Deep learning has led to dramatic performance improvements for the task of speech enhancement, where deep neural networks (DNNs) are trained to recover clean speech from noisy and reverberant mixtures. Most of the existing DNN-based algorithms operate in the frequency domain, as time-domain approaches are believed to be less effective for speech dereverberation. In this study, we employ two DNNs: ARN (attentive recurrent network) and DC-CRN (densely-connected convolutional recurrent network), and systematically investigate the effects of different components on enhancement performance, such as window sizes, loss functions, and feature representations. We conduct evaluation experiments in two main conditions: reverberant-only and reverberant-noisy. Our findings suggest that incorporating larger window sizes is helpful for dereverberation, and adding transform operations (either convolutional or linear) to encode and decode waveform features improves the sparsity of the learned representations, and boosts the performance of time-domain models. Experimental results demonstrate that ARN and DC-CRN with proposed techniques achieve superior performance compared with other strong enhancement baselines.

Speech signals are valuable biomarkers for assessing an individual's mental health, including identifying Major Depressive Disorder (MDD) automatically. A frequently used approach in this regard is to employ features related to speaker identity, such as speaker-embeddings. However, over-reliance on speaker identity features in mental health screening systems can compromise patient privacy. Moreover, some aspects of speaker identity may not be relevant for depression detection and could serve as a bias factor that hampers system performance. To overcome these limitations, we propose disentangling speaker-identity information from depression-related information. Specifically, we present four distinct disentanglement methods to achieve this - adversarial speaker identification (SID)-loss maximization (ADV), SID-loss equalization with variance (LEV), SID-loss equalization using Cross-Entropy (LECE) and SID-loss equalization using KL divergence (LEKLD). Our experiments, which incorporated diverse input features and model architectures, have yielded improved F1 scores for MDD detection and voice-privacy attributes, as quantified by Gain in Voice Distinctiveness and De-Identification Scores (DeID). On the DAIC-WOZ dataset (English), LECE using ComparE16 features results in the best F1-Scores of 80% which represents the audio-only SOTA depression detection F1-Score along with a of -1.1 dB and a DeID of 85%. On the EATD dataset (Mandarin), ADV using raw-audio signal achieves an F1-Score of 72.38% surpassing multi-modal SOTA along with a of -0.89 dB dB and a DeID of 51.21%. By reducing the dependence on speaker-identity-related features, our method offers a promising direction for speech-based depression detection that preserves patient privacy.

Text-based computational approaches for assessing the quality of psychotherapy are being developed to support quality assurance and clinical training. However, due to the long durations of typical conversation based therapy sessions, and due to limited annotated modeling resources, computational methods largely rely on frequency-based lexical features or dialogue acts to assess the overall session level characteristics. In this work, we propose a hierarchical framework to automatically evaluate the quality of transcribed Cognitive Behavioral Therapy (CBT) interactions. Given the richly dynamic nature of the spoken dialog within a talk therapy session, to evaluate the overall session level quality, we propose to consider modeling it as a function of local variations across the interaction. To implement that empirically, we divide each psychotherapy session into conversation segments and initialize the segment-level qualities with the session-level scores. First, we produce segment embeddings by fine-tuning a BERT-based model, and predict segment-level (local) quality scores. These embeddings are used as the lower-level input to a Bidirectional LSTM-based neural network to predict the session-level (global) quality estimates. In particular, we model the global quality as a linear function of the local quality scores, which allows us to update the segment-level quality estimates based on the session-level quality prediction. These newly estimated segment-level scores benefit the BERT fine-tuning process, which in turn results in better segment embeddings. We evaluate the proposed framework on automatically derived transcriptions from real-world CBT clinical recordings to predict session-level behavior codes. The results indicate that our approach leads to improved evaluation accuracy for most codes when used for both regression and classification tasks.

The technology development for point-of-care tests (POCTs) targeting respiratory diseases has witnessed a growing demand in the recent past. Investigating the presence of acoustic biomarkers in modalities such as cough, breathing and speech sounds, and using them for building POCTs can offer fast, contactless and inexpensive testing. In view of this, over the past year, we launched the "Coswara" project to collect cough, breathing and speech sound recordings via worldwide crowdsourcing. With this data, a call for development of diagnostic tools was announced in the Interspeech 2021 as a special session titled "Diagnostics of COVID-19 using Acoustics (DiCOVA) Challenge". The goal was to bring together researchers and practitioners interested in developing acoustics-based COVID-19 POCTs by enabling them to work on the same set of development and test datasets. As part of the challenge, datasets with breathing, cough, and speech sound samples from COVID-19 and non-COVID-19 individuals were released to the participants. The challenge consisted of two tracks. The Track-1 focused only on cough sounds, and participants competed in a leaderboard setting. In Track-2, breathing and speech samples were provided for the participants, without a competitive leaderboard. The challenge attracted 85 plus registrations with 29 final submissions for Track-1. This paper describes the challenge (datasets, tasks, baseline system), and presents a focused summary of the various systems submitted by the participating teams. An analysis of the results from the top four teams showed that a fusion of the scores from these teams yields an area-under-the-receiver operating curve (AUC-ROC) of 95.1% on the blind test data. By summarizing the lessons learned, we foresee the challenge overview in this paper to help accelerate technological development of acoustic-based POCTs.

In a conversational exchange, interlocutors use social cues including conversational turn-taking to communicate. There has been attention in the literature concerning how mothers, fathers, boys, and girls converse with each other, and in particular who initiates a conversation. Better understanding of conversational dynamics may deepen our understanding of social roles, speech and language development, and individual language variability. Here we use large-scale automatic analysis of 186 naturalistic daylong acoustic recordings to examine the conversational dynamics of 26 families with children about 30 months of age to better understand communication roles. Families included 15 with boys and 11 with girls. There was no difference in conversation initiation rate by child sex, but children initiated more conversations than mothers, and mothers initiated more than fathers. Results support developmental theories of the different and variable roles that interlocutors play in a social context.

Early detection of cognitive decline involved in Alzheimer's Disease and Related Dementias (ADRD) in older adults living alone is essential for developing, planning, and initiating interventions and support systems to improve users' everyday function and quality of life. In this paper, we explore the voice commands using a Voice-Assistant System (VAS), i.e., Amazon Alexa, from 40 older adults who were either Healthy Control (HC) participants or Mild Cognitive Impairment (MCI) participants, age 65 or older. We evaluated the data collected from voice commands, cognitive assessments, and interviews and surveys using a structured protocol. We extracted 163 unique command-relevant features from each participant's use of the VAS. We then built machine-learning models including 1-layer/2-layer neural networks, support vector machines, decision tree, and random forest, for classification and comparison with standard cognitive assessment scores, e.g., Montreal Cognitive Assessment (MoCA). Our classification models using fusion features achieved an accuracy of 68%, and our regression model resulted in a Root-Mean-Square Error (RMSE) score of 3.53. Our Decision Tree (DT) and Random Forest (RF) models using selected features achieved higher classification accuracy 80-90%. Finally, we analyzed the contribution of each feature set to the model output, thus revealing the commands and features most useful in inferring the participants' cognitive status. We found that features of overall performance, features of music-related commands, features of call-related commands, and features from Automatic Speech Recognition (ASR) were the top-four feature sets most impactful on inference accuracy. The results from this controlled study demonstrate the promise of future home-based cognitive assessments using Voice-Assistant Systems.

When interviewing a child who may have witnessed a crime, the interviewer must ask carefully directed questions in order to elicit a truthful statement from the child. The presented work uses Granger causal analysis to examine and represent child-interviewer interaction dynamics over such an interview. Our work demonstrates that Granger Causal analysis of psycholinguistic and acoustic signals from speech yields significant predictors of whether a child is telling the truth, as well as whether a child will disclose witnessing a transgression later in the interview. By incorporating cross-modal Granger causal features extracted from audio and transcripts of forensic interviews, we are able to substantially outperform conventional deception detection methods and a number of simulated baselines. Our results suggest that a child's use of concreteness and imageability in their language are strong psycholinguistic indicators of truth-telling and that the coordination of child and interviewer speech signals is much more informative than the specific language used throughout the interview.

Anomia (word-finding difficulties) is the hallmark of aphasia, an acquired language disorder most commonly caused by stroke. Assessment of speech performance using picture naming tasks is a key method for both diagnosis and monitoring of responses to treatment interventions by people with aphasia (PWA). Currently, this assessment is conducted manually by speech and language therapists (SLT). Surprisingly, despite advancements in automatic speech recognition (ASR) and artificial intelligence with technologies like deep learning, research on developing automated systems for this task has been scarce. Here we present NUVA, an utterance verification system incorporating a deep learning element that classifies 'correct' versus' incorrect' naming attempts from aphasic stroke patients. When tested on eight native British-English speaking PWA the system's performance accuracy ranged between 83.6% to 93.6%, with a 10-fold cross-validation mean of 89.5%. This performance was not only significantly better than a baseline created for this study using one of the leading commercially available ASRs (Google speech-to-text service) but also comparable in some instances with two independent SLT ratings for the same dataset.

Emotional speech production has been previously studied using fleshpoint tracking data in speaker-specific experiment setups. The present study introduces a real-time magnetic resonance imaging database of emotional speech production from 10 speakers and presents articulatory analysis results of speech emotional expression using the database. Midsagittal vocal tract parameters (midsagittal distances and the vocal tract length) were parameterized based on a two-dimensional grid-line system, using image segmentation software. The principal feature analysis technique was applied to the grid-line system in order to find the major movement locations. Results reveal both speaker-dependent and speaker-independent variation patterns. For example, sad speech, a low arousal emotion, tends to show smaller opening for low vowels in the front cavity than the high arousal emotions more consistently than the other regions of the vocal tract. Happiness shows significantly shorter vocal tract length than anger and sadness in most speakers. Further details of speaker-dependent and speaker-independent speech articulation variation in emotional expression and their implications are described.

Automatic speech recognition for child speech has been long considered a more challenging problem than for adult speech. Various contributing factors have been identified such as larger acoustic speech variability including mispronunciations due to continuing biological changes in growth, developing vocabulary and linguistic skills, and scarcity of training corpora. A further challenge arises when dealing with spontaneous speech of children involved in a conversational interaction, and especially when the child may have limited or impaired communication ability. This includes health applications, one of the motivating domains of this paper, that involve goal-oriented dyadic interactions between a child and clinician/adult social partner as a part of behavioral assessment. In this work, we use linguistic context information from the interaction to adapt speech recognition models for children speech. Specifically, spoken language from the interacting adult speech provides the context for the child's speech. We propose two methods to exploit this context: lexical repetitions and semantic response generation. For the latter, we make use of sequence-to-sequence models that learn to predict the target child utterance given context adult utterances. Long-term context is incorporated in the model by propagating the cell-state across the duration of conversation. We use interpolation techniques to adapt language models at the utterance level, and analyze the effect of length and direction of context (forward and backward). Two different domains are used in our experiments to demonstrate the generalized nature of our methods - interactions between a child with ASD and an adult social partner in a play-based, naturalistic setting, and in forensic interviews between a child and a trained interviewer. In both cases, context-adapted models yield significant improvement (upto 10.71% in absolute word error rate) over the baseline and perform consistently across context windows and directions. Using statistical analysis, we investigate the effect of source-based (adult) and target-based (child) factors on adaptation methods. Our results demonstrate the applicability of our modeling approach in improving child speech recognition by employing information transfer from the adult interlocutor.

Children speech recognition is challenging mainly due to the inherent high variability in children's physical and articulatory characteristics and expressions. This variability manifests in both acoustic constructs and linguistic usage due to the rapidly changing developmental stage in children's life. Part of the challenge is due to the lack of large amounts of available children speech data for efficient modeling. This work attempts to address the key challenges using transfer learning from adult's models to children's models in a Deep Neural Network (DNN) framework for children's Automatic Speech Recognition (ASR) task evaluating on multiple children's speech corpora with a large vocabulary. The paper presents a systematic and an extensive analysis of the proposed transfer learning technique considering the key factors affecting children's speech recognition from prior literature. are presented on (i) comparisons of earlier GMM-HMM and the newer DNN Models, (ii) effectiveness of standard adaptation techniques versus transfer learning, (iii) various adaptation configurations in tackling the variabilities present in children speech, in terms of (a) acoustic spectral variability, and (b) pronunciation variability and linguistic constraints. Our spans over (i) number of DNN model parameters (for adaptation), (ii) amount of adaptation data, (iii) ages of children, (iv) age dependent-independent adaptation. Finally, we provide on (i) the favorable strategies over various aforementioned - analyzed parameters, and (ii) potential future research directions and relevant challenges/problems persisting in DNN based ASR for children's speech.

The National Institute of Standards and Technology has been conducting Speaker Recognition Evaluations (SREs) for over 20 years. This article provides an overview of the practice of evaluating speaker recognition technology as it has evolved during this time. Focus is given to the current state of speaker recognition evaluation. Highlights from past SREs and future plans are also discussed.

We introduce a real-time capable algorithm which estimates the long-term signal to noise ratio (SNR) of the speech in multi-talker babble noise. In real-time applications, long-term SNR is calculated over a sufficiently long moving frame of the noisy speech ending at the current time. The algorithm performs the real-time long-term SNR estimation by averaging "speech-likeness" values of multiple consecutive short-frames of the noisy speech which collectively form a long-frame with an adaptive length. The algorithm is calibrated to be insensitive to short-term fluctuations and transient changes in speech or noise level. However, it quickly responds to non-transient changes in long-term SNR by adjusting the duration of the long-frame on which the long-term SNR is measured. This ability is obtained by employing an event detector and adaptive frame duration. The event detector identifies non-transient changes of the long-term SNR and optimizes the duration of the long-frame accordingly. The algorithm was trained and tested for randomly generated speech samples corrupted with multi-talker babble. In addition to its ability to provide an adaptive long-term SNR estimation in a dynamic noisy situation, the evaluation results show that the algorithm outperforms the existing overall SNR estimation methods in multi-talker babble over a wide range of number of talkers and SNRs. The relatively low computational cost and the ability to update the estimated long-term SNR several times per second make this algorithm capable of operating in real-time speech processing applications.

Computer-Assisted Pronunciation Training (CAPT) systems aim to help a child learn the correct pronunciations of words. However, while there are many online commercial CAPT apps, there is no consensus among Speech Language Therapists (SLPs) or non-professionals about which CAPT systems, if any, work well. The prevailing assumption is that practicing with such programs is less reliable and thus does not provide the feedback necessary to allow children to improve their performance. The most common method for assessing pronunciation performance is the Goodness of Pronunciation (GOP) technique. Our paper proposes two new GOP techniques. We have found that pronunciation models that use explicit knowledge about error pronunciation patterns can lead to more accurate classification whether a phoneme was correctly pronounced or not. We evaluate the proposed pronunciation assessment methods against a baseline state of the art GOP approach, and show that the proposed techniques lead to classification performance that is more similar to that of a human expert.

Methods from automatic speech recognition (ASR), such as segmentation and forced alignment, have facilitated the rapid annotation and analysis of very large adult speech databases and databases of caregiver-infant interaction, enabling advances in speech science that were unimaginable just a few decades ago. This paper centers on two main problems that must be addressed in order to have analogous resources for developing and exploiting databases of young children's speech. The first problem is to understand and appreciate the differences between adult and child speech that cause ASR models developed for adult speech to fail when applied to child speech. These differences include the fact that children's vocal tracts are smaller than those of adult males and also changing rapidly in size and shape over the course of development, leading to between-talker variability across age groups that dwarfs the between-talker differences between adult men and women. Moreover, children do not achieve fully adult-like speech motor control until they are young adults, and their vocabularies and phonological proficiency are developing as well, leading to considerably more within-talker variability as well as more between-talker variability. The second problem then is to determine what annotation schemas and analysis techniques can most usefully capture relevant aspects of this variability. Indeed, standard acoustic characterizations applied to child speech reveal that adult-centered annotation schemas fail to capture phenomena such as the emergence of covert contrasts in children's developing phonological systems, while also revealing children's nonuniform progression toward community speech norms as they acquire the phonological systems of their native languages. Both problems point to the need for more basic research into the growth and development of the articulatory system (as well as of the lexicon and phonological system) that is oriented explicitly toward the construction of age-appropriate computational models.

Child engagement is defined as the interaction of a child with his/her environment in a contextually appropriate manner. Engagement behavior in children is linked to socio-emotional and cognitive state assessment with enhanced engagement identified with improved skills. A vast majority of studies however rely solely, and often implicitly, on subjective perceptual measures of engagement. Access to automatic quantification could assist researchers/clinicians to objectively interpret engagement with respect to a target behavior or condition, and furthermore inform mechanisms for improving engagement in various settings. In this paper, we present an engagement prediction system based exclusively on vocal cues observed during structured interaction between a child and a psychologist involving several tasks. Specifically, we derive prosodic cues that capture engagement levels across the various tasks. Our experiments suggest that a child's engagement is reflected not only in the vocalizations, but also in the speech of the interacting psychologist. Moreover, we show that prosodic cues are informative of the engagement phenomena not only as characterized over the entire task (i.e., global cues), but also in short term patterns (i.e., local cues). We perform a classification experiment assigning the engagement of a child into three discrete levels achieving an unweighted average recall of 55.8% (chance is 33.3%). While the systems using global cues and local level cues are each statistically significant in predicting engagement, we obtain the best results after fusing these two components. We perform further analysis of the cues at local and global levels to achieve insights linking specific prosodic patterns to the engagement phenomenon. We observe that while the performance of our model varies with task setting and interacting psychologist, there exist universal prosodic patterns reflective of engagement.

How the speech production and perception systems evolved in humans still remains a mystery today. Previous research suggests that human auditory systems are able, and have possibly evolved, to preserve maximal information about the speaker's articulatory gestures. This paper attempts an initial step towards answering the question of whether speakers' articulatory mechanisms have also evolved to produce sounds that can be optimally discriminated by the listener's auditory system. To this end we explicitly model, using computational methods, the extent to which derived representations of "primitive movements" of speech articulation can be used to discriminate between broad phone categories. We extract spatio-temporal primitive movements as recurring patterns in a data matrix of human speech articulation, i.e. representing the trajectories of vocal tract articulators over time. To this end, we propose a weakly-supervised learning method that attempts to find a part-based representation of the data in terms of recurring basis trajectory units (or primitives) and their corresponding activations over time. For each phone interval, we then derive a feature representation that captures the co-occurrences between the activations of the various bases over different time-lags. We show that this feature, derived entirely from activations of these primitive movements, is able to achieve a greater discrimination relative to using conventional features on an interval-based phone classification task. We discuss the implications of these findings in furthering our understanding of speech signal representations and the links between speech production and perception systems.

We propose a , feature-level and score-level fusion approach by combining acoustic and estimated articulatory information for both text independent and text dependent speaker verification. From a practical point of view, we study how to improve speaker verification performance by combining dynamic articulatory information with the conventional acoustic features. On text independent speaker verification, we find that concatenating articulatory features obtained from measured speech production data with conventional Mel-frequency cepstral coefficients (MFCCs) improves the performance dramatically. However, since directly measuring articulatory data is not feasible in many real world applications, we also experiment with estimated articulatory features obtained through acoustic-to-articulatory inversion. We explore both feature level and score level fusion methods and find that the overall system performance is significantly enhanced even with estimated articulatory features. Such a performance boost could be due to the inter-speaker variation information embedded in the estimated articulatory features. Since the dynamics of articulation contain important information, we included inverted articulatory trajectories in text dependent speaker verification. We demonstrate that the articulatory constraints introduced by inverted articulatory features help to reject wrong password trials and improve the performance after score level fusion. We evaluate the proposed methods on the X-ray Microbeam database and the RSR 2015 database, respectively, for the aforementioned two tasks. Experimental results show that we achieve more than 15% relative equal error rate reduction for both speaker verification tasks.

Non-verbal communication involves encoding, transmission and decoding of non-lexical cues and is realized using vocal (e.g. prosody) or visual (e.g. gaze, body language) channels during conversation. These cues perform the function of maintaining conversational flow, expressing emotions, and marking personality and interpersonal attitude. In particular, non-verbal cues in speech such as paralanguage and non-verbal vocal events (e.g. laughters, sighs, cries) are used to nuance meaning and convey emotions, mood and attitude. For instance, laughters are associated with affective expressions while fillers (e.g. um, ah, um) are used to hold floor during a conversation. In this paper we present an automatic non-verbal vocal events detection system focusing on the detect of laughter and fillers. We extend our system presented during Interspeech 2013 Social Signals Sub-challenge (that was the winning entry in the challenge) for frame-wise event detection and test several schemes for incorporating local context during detection. Specifically, we incorporate context at two separate levels in our system: (i) the raw frame-wise features and, (ii) the output decisions. Furthermore, our system processes the output probabilities based on a few heuristic rules in order to reduce erroneous frame-based predictions. Our overall system achieves an Area Under the Receiver Operating Characteristics curve of 95.3% for detecting laughters and 90.4% for fillers on the test set drawn from the data specifications of the Interspeech 2013 Social Signals Sub-challenge. We perform further analysis to understand the interrelation between the features and obtained results. Specifically, we conduct a feature sensitivity analysis and correlate it with each feature's stand alone performance. The observations suggest that the trained system is more sensitive to a feature carrying higher discriminability with implications towards a better system design.

Pathological speech usually refers to the condition of speech distortion resulting from atypicalities in voice and/or in the articulatory mechanisms owing to disease, illness or other physical or biological insult to the production system. Although automatic evaluation of speech intelligibility and quality could come in handy in these scenarios to assist experts in diagnosis and treatment design, the many sources and types of variability often make it a very challenging computational processing problem. In this work we propose novel sentence-level features to capture abnormal variation in the prosodic, voice quality and pronunciation aspects in pathological speech. In addition, we propose a post-classification posterior smoothing scheme which refines the posterior of a test sample based on the posteriors of other test samples. Finally, we perform feature-level fusions and subsystem decision fusion for arriving at a final intelligibility decision. The performances are tested on two pathological speech datasets, the NKI CCRT Speech Corpus (advanced head and neck cancer) and the TORGO database (cerebral palsy or amyotrophic lateral sclerosis), by evaluating classification accuracy without overlapping subjects' data among training and test partitions. Results show that the feature sets of each of the voice quality subsystem, prosodic subsystem, and pronunciation subsystem, offer significant discriminating power for binary intelligibility classification. We observe that the proposed posterior smoothing in the acoustic space can further reduce classification errors. The smoothed posterior score fusion of subsystems shows the best classification performance (73.5% for unweighted, and 72.8% for weighted, average recalls of the binary classes).

We introduce a ranking approach for emotion recognition which naturally incorporates information about the general expressivity of speakers. We demonstrate that our approach leads to substantial gains in accuracy compared to conventional approaches. We train ranking SVMs for individual emotions, treating the data from each speaker as a separate query, and combine the predictions from all rankers to perform multi-class prediction. The ranking method provides two natural benefits. It captures speaker specific information even in speaker-independent training/testing conditions. It also incorporates the intuition that each utterance can express a mix of possible emotion and that considering the degree to which each emotion is expressed can be productively exploited to identify the dominant emotion. We compare the performance of the rankers and their combination to standard SVM classification approaches on two publicly available datasets of acted emotional speech, Berlin and LDC, as well as on spontaneous emotional data from the FAU Aibo dataset. On acted data, ranking approaches exhibit significantly better performance compared to SVM classification both in distinguishing a specific emotion from all others and in multi-class prediction. On the spontaneous data, which contains mostly neutral utterances with a relatively small portion of less intense emotional utterances, ranking-based classifiers again achieve much higher precision in identifying emotional utterances than conventional SVM classifiers. In addition, we discuss the complementarity of conventional SVM and ranking-based classifiers. On all three datasets we find dramatically higher accuracy for the test items on whose prediction the two methods agree compared to the accuracy of individual methods. Furthermore on the spontaneous data the ranking and standard classification are complementary and we obtain marked improvement when we combine the two classifiers by late-stage fusion.