Carol Espy-Wilson

Rahil Parikh, Ilya Kavalerov, Carol Espy-Wilson et al. · amazon-science

Recent advancements in deep learning have led to drastic improvements in speech segregation models. Despite their success and growing applicability, few efforts have been made to analyze the underlying principles that these networks learn to perform segregation. Here we analyze the role of harmonicity on two state-of-the-art Deep Neural Networks (DNN)-based models- Conv-TasNet and DPT-Net. We evaluate their performance with mixtures of natural speech versus slightly manipulated inharmonic speech, where harmonics are slightly frequency jittered. We find that performance deteriorates significantly if one source is even slightly harmonically jittered, e.g., an imperceptible 3% harmonic jitter degrades performance of Conv-TasNet from 15.4 dB to 0.70 dB. Training the model on inharmonic speech does not remedy this sensitivity, instead resulting in worse performance on natural speech mixtures, making inharmonicity a powerful adversarial factor in DNN models. Furthermore, additional analyses reveal that DNN algorithms deviate markedly from biologically inspired algorithms that rely primarily on timing cues and not harmonicity to segregate speech.

Rahil Parikh, Nadee Seneviratne, Ganesh Sivaraman et al. · amazon-science

Multi-resolution spectro-temporal features of a speech signal represent how the brain perceives sounds by tuning cortical cells to different spectral and temporal modulations. These features produce a higher dimensional representation of the speech signals. The purpose of this paper is to evaluate how well the auditory cortex representation of speech signals contribute to estimate articulatory features of those corresponding signals. Since obtaining articulatory features from acoustic features of speech signals has been a challenging topic of interest for different speech communities, we investigate the possibility of using this multi-resolution representation of speech signals as acoustic features. We used U. of Wisconsin X-ray Microbeam (XRMB) database of clean speech signals to train a feed-forward deep neural network (DNN) to estimate articulatory trajectories of six tract variables. The optimal set of multi-resolution spectro-temporal features to train the model were chosen using appropriate scale and rate vector parameters to obtain the best performing model. Experiments achieved a correlation of 0.675 with ground-truth tract variables. We compared the performance of this speech inversion system with prior experiments conducted using Mel Frequency Cepstral Coefficients (MFCCs).

Rahil Parikh, Gaspar Rochette, Carol Espy-Wilson et al. · amazon-science

End-to-end learning models have demonstrated a remarkable capability in performing speech segregation. Despite their wide-scope of real-world applications, little is known about the mechanisms they employ to group and consequently segregate individual speakers. Knowing that harmonicity is a critical cue for these networks to group sources, in this work, we perform a thorough investigation on ConvTasnet and DPT-Net to analyze how they perform a harmonic analysis of the input mixture. We perform ablation studies where we apply low-pass, high-pass, and band-stop filters of varying pass-bands to empirically analyze the harmonics most critical for segregation. We also investigate how these networks decide which output channel to assign to an estimated source by introducing discontinuities in synthetic mixtures. We find that end-to-end networks are highly unstable, and perform poorly when confronted with deformations which are imperceptible to humans. Replacing the encoder in these networks with a spectrogram leads to lower overall performance, but much higher stability. This work helps us to understand what information these network rely on for speech segregation, and exposes two sources of generalization-errors. It also pinpoints the encoder as the part of the network responsible for these errors, allowing for a redesign with expert knowledge or transfer learning.

Ahmed Adel Attia, Carol Espy-Wilson

Articulatory recordings track the positions and motion of different articulators along the vocal tract and are widely used to study speech production and to develop speech technologies such as articulatory based speech synthesizers and speech inversion systems. The University of Wisconsin X-Ray microbeam (XRMB) dataset is one of various datasets that provide articulatory recordings synced with audio recordings. The XRMB articulatory recordings employ pellets placed on a number of articulators which can be tracked by the microbeam. However, a significant portion of the articulatory recordings are mistracked, and have been so far unsuable. In this work, we present a deep learning based approach using Masked Autoencoders to accurately reconstruct the mistracked articulatory recordings for 41 out of 47 speakers of the XRMB dataset. Our model is able to reconstruct articulatory trajectories that closely match ground truth, even when three out of eight articulators are mistracked, and retrieve 3.28 out of 3.4 hours of previously unusable recordings.

Ahmed Adel Attia, Jing Liu, Wei Ai et al.

Recent advancements in Automatic Speech Recognition (ASR) systems, exemplified by Whisper, have demonstrated the potential of these systems to approach human-level performance given sufficient data. However, this progress doesn't readily extend to ASR for children due to the limited availability of suitable child-specific databases and the distinct characteristics of children's speech. A recent study investigated leveraging the My Science Tutor (MyST) children's speech corpus to enhance Whisper's performance in recognizing children's speech. They were able to demonstrate some improvement on a limited testset. This paper builds on these findings by enhancing the utility of the MyST dataset through more efficient data preprocessing. We reduce the Word Error Rate (WER) on the MyST testset 13.93% to 9.11% with Whisper-Small and from 13.23% to 8.61% with Whisper-Medium and show that this improvement can be generalized to unseen datasets. We also highlight important challenges towards improving children's ASR performance. The results showcase the viable and efficient integration of Whisper for effective children's speech recognition.

Ahmed Adel Attia, Yashish M. Siriwardena, Carol Espy-Wilson

The performance of deep learning models depends significantly on their capacity to encode input features efficiently and decode them into meaningful outputs. Better input and output representation has the potential to boost models' performance and generalization. In the context of acoustic-to-articulatory speech inversion (SI) systems, we study the impact of utilizing speech representations acquired via self-supervised learning (SSL) models, such as HuBERT compared to conventional acoustic features. Additionally, we investigate the incorporation of novel tract variables (TVs) through an improved geometric transformation model. By combining these two approaches, we improve the Pearson product-moment correlation (PPMC) scores which evaluate the accuracy of TV estimation of the SI system from 0.7452 to 0.8141, a 6.9% increase. Our findings underscore the profound influence of rich feature representations from SSL models and improved geometric transformations with target TVs on the enhanced functionality of SI systems.

Ahmed Adel Attia, Dorottya Demszky, Tolulope Ogunremi et al.

Creating Automatic Speech Recognition (ASR) systems that are robust and resilient to classroom conditions is paramount to the development of AI tools to aid teachers and students. In this work, we study the efficacy of continued pretraining (CPT) in adapting Wav2vec2.0 to the classroom domain. We show that CPT is a powerful tool in that regard and reduces the Word Error Rate (WER) of Wav2vec2.0-based models by upwards of 10%. More specifically, CPT improves the model's robustness to different noises, microphones and classroom conditions.

Gowtham Premananth, Carol Espy-Wilson

Language disruptions are one of the well-known effects of schizophrenia symptoms. They are often manifested as disorganized speech and impaired discourse coherence. These abnormalities in spontaneous language production reflect underlying cognitive disturbances and have the potential to serve as objective markers for symptom severity and diagnosis of schizophrenia. This study focuses on how these language disruptions can be characterized in terms of two computational linguistic measures: surprisal and semantic coherence. By computing surprisal and semantic coherence of language using computational models, this study investigates how they differ between subjects with schizophrenia and healthy controls. Furthermore, this study provides further insight into how language disruptions in terms of these linguistic measures change with varying degrees of schizophrenia symptom severity.

Gowtham Premananth, Carol Espy-Wilson

Advances in artificial intelligence (AI) and deep learning have improved diagnostic capabilities in healthcare, yet limited interpretability continues to hinder clinical adoption. Schizophrenia, a complex disorder with diverse symptoms including disorganized speech and social withdrawal, demands tools that capture symptom severity and provide clinically meaningful insights beyond binary diagnosis. Here, we present an interpretable framework that leverages articulatory speech features through eigenspectra difference plots and a weighted sum with exponential decay (WSED) to quantify vocal tract coordination. Eigenspectra plots effectively distinguished complex from simpler coordination patterns, and WSED scores reliably separated these groups, with ambiguity confined to a narrow range near zero. Importantly, WSED scores correlated not only with overall BPRS severity but also with the balance between positive and negative symptoms, reflecting more complex coordination in subjects with pronounced positive symptoms and the opposite trend for stronger negative symptoms. This approach offers a transparent, severity-sensitive biomarker for schizophrenia, advancing the potential for clinically interpretable speech-based assessment tools.

Thanushi Withanage, Elizabeth Redcay, Carol Espy-Wilson

Individuals often align their speaking patterns with their interlocutors, a phenomenon linked to engagement and rapport. While well documented in task-oriented dialogues, less is known about entrainment in naturalistic, non-task and virtual settings. In this study, we analyze a large corpus of spontaneous dyadic Zoom conversations to examine how conversational dynamics relate to perceived interaction quality. We extract multimodal features encompassing turn-taking, pauses, facial movements, and acoustic measures such as pitch and intensity. Perceived conversational success was quantified via factor analysis of post-conversation ratings. Results demonstrate that entrainment reliably detected in spontaneous speech and correlates with higher perceived success. These findings identify key interactional markers of conversational quality and highlight opportunities for targeted interventions to foster more effective and engaging communication.

Gowtham Premananth, Philip Resnik, Sonia Bansal et al.

Studies on schizophrenia assessments using deep learning typically treat it as a classification task to detect the presence or absence of the disorder, oversimplifying the condition and reducing its clinical applicability. This traditional approach overlooks the complexity of schizophrenia, limiting its practical value in healthcare settings. This study shifts the focus to individual symptom severity estimation using a multimodal approach that integrates speech, video, and text inputs. We develop unimodal models for each modality and a multimodal framework to improve accuracy and robustness. By capturing a more detailed symptom profile, this approach can help in enhancing diagnostic precision and support personalized treatment, offering a scalable and objective tool for mental health assessment.

Ahmed Adel Attia, Dorottya Demszky, Tolulope Ogunremi et al.

Creating Automatic Speech Recognition (ASR) systems that are robust and resilient to classroom conditions is paramount to the development of AI tools to aid teachers and students. In this work, we study the efficacy of continued pretraining (CPT) in adapting Wav2vec2.0 to the classroom domain. We show that CPT is a powerful tool in that regard and reduces the Word Error Rate (WER) of Wav2vec2.0-based models by upwards of 10%. More specifically, CPT improves the model's robustness to different noises, microphones, classroom conditions as well as classroom demographics. Our CPT models show improved ability to generalize to different demographics unseen in the labeled finetuning data.

Ahmed Adel Attia, Dorottya Demszky, Jing Liu et al.

Recent progress in speech recognition has relied on models trained on vast amounts of labeled data. However, classroom Automatic Speech Recognition (ASR) faces the real-world challenge of abundant weak transcripts paired with only a small amount of accurate, gold-standard data. In such low-resource settings, high transcription costs make re-transcription impractical. To address this, we ask: what is the best approach when abundant inexpensive weak transcripts coexist with limited gold-standard data, as is the case for classroom speech data? We propose Weakly Supervised Pretraining (WSP), a two-step process where models are first pretrained on weak transcripts in a supervised manner, and then fine-tuned on accurate data. Our results, based on both synthetic and real weak transcripts, show that WSP outperforms alternative methods, establishing it as an effective training methodology for low-resource ASR in real-world scenarios.

Nadee Seneviratne, Carol Espy-Wilson

Multimodal depression classification has gained immense popularity over the recent years. We develop a multimodal depression classification system using articulatory coordination features extracted from vocal tract variables and text transcriptions obtained from an automatic speech recognition tool that yields improvements of area under the receiver operating characteristics curve compared to uni-modal classifiers (7.5% and 13.7% for audio and text respectively). We show that in the case of limited training data, a segment-level classifier can first be trained to then obtain a session-wise prediction without hindering the performance, using a multi-stage convolutional recurrent neural network. A text model is trained using a Hierarchical Attention Network (HAN). The multimodal system is developed by combining embeddings from the session-level audio model and the HAN text model

Yashish M. Siriwardena, Chris Kitchen, Deanna L. Kelly et al.

This study investigates the speech articulatory coordination in schizophrenia subjects exhibiting strong positive symptoms (e.g. hallucinations and delusions), using two distinct channel-delay correlation methods. We show that the schizophrenic subjects with strong positive symptoms and who are markedly ill pose complex articulatory coordination pattern in facial and speech gestures than what is observed in healthy subjects. This distinction in speech coordination pattern is used to train a multimodal convolutional neural network (CNN) which uses video and audio data during speech to distinguish schizophrenic patients with strong positive symptoms from healthy subjects. We also show that the vocal tract variables (TVs) which correspond to place of articulation and glottal source outperform the Mel-frequency Cepstral Coefficients (MFCCs) when fused with Facial Action Units (FAUs) in the proposed multimodal network. For the clinical dataset we collected, our best performing multimodal network improves the mean F1 score for detecting schizophrenia by around 18% with respect to the full vocal tract coordination (FVTC) baseline method implemented with fusing FAUs and MFCCs.

Nadee Seneviratne, Carol Espy-Wilson

Speech based depression classification has gained immense popularity over the recent years. However, most of the classification studies have focused on binary classification to distinguish depressed subjects from non-depressed subjects. In this paper, we formulate the depression classification task as a severity level classification problem to provide more granularity to the classification outcomes. We use articulatory coordination features (ACFs) developed to capture the changes of neuromotor coordination that happens as a result of psychomotor slowing, a necessary feature of Major Depressive Disorder. The ACFs derived from the vocal tract variables (TVs) are used to train a dilated Convolutional Neural Network based depression classification model to obtain segment-level predictions. Then, we propose a Recurrent Neural Network based approach to obtain session-level predictions from segment-level predictions. We show that strengths of the segment-wise classifier are amplified when a session-wise classifier is trained on embeddings obtained from it. The model trained on ACFs derived from TVs show relative improvement of 27.47% in Unweighted Average Recall (UAR) at the session-level classification task, compared to the ACFs derived from Mel Frequency Cepstral Coefficients (MFCCs).

Nadee Seneviratne, Carol Espy-Wilson

Depression detection using vocal biomarkers is a highly researched area. Articulatory coordination features (ACFs) are developed based on the changes in neuromotor coordination due to psychomotor slowing, a key feature of Major Depressive Disorder. However findings of existing studies are mostly validated on a single database which limits the generalizability of results. Variability across different depression databases adversely affects the results in cross corpus evaluations (CCEs). We propose to develop a generalized classifier for depression detection using a dilated Convolutional Neural Network which is trained on ACFs extracted from two depression databases. We show that ACFs derived from Vocal Tract Variables (TVs) show promise as a robust set of features for depression detection. Our model achieves relative accuracy improvements of ~10% compared to CCEs performed on models trained on a single database. We extend the study to show that fusing TVs and Mel-Frequency Cepstral Coefficients can further improve the performance of this classifier.

Matthew Marge, Carol Espy-Wilson, Nigel Ward

With robotics rapidly advancing, more effective human-robot interaction is increasingly needed to realize the full potential of robots for society. While spoken language must be part of the solution, our ability to provide spoken language interaction capabilities is still very limited. The National Science Foundation accordingly convened a workshop, bringing together speech, language, and robotics researchers to discuss what needs to be done. The result is this report, in which we identify key scientific and engineering advances needed. Our recommendations broadly relate to eight general themes. First, meeting human needs requires addressing new challenges in speech technology and user experience design. Second, this requires better models of the social and interactive aspects of language use. Third, for robustness, robots need higher-bandwidth communication with users and better handling of uncertainty, including simultaneous consideration of multiple hypotheses and goals. Fourth, more powerful adaptation methods are needed, to enable robots to communicate in new environments, for new tasks, and with diverse user populations, without extensive re-engineering or the collection of massive training data. Fifth, since robots are embodied, speech should function together with other communication modalities, such as gaze, gesture, posture, and motion. Sixth, since robots operate in complex environments, speech components need access to rich yet efficient representations of what the robot knows about objects, locations, noise sources, the user, and other humans. Seventh, since robots operate in real time, their speech and language processing components must also. Eighth, in addition to more research, we need more work on infrastructure and resources, including shareable software modules and internal interfaces, inexpensive hardware, baseline systems, and diverse corpora.

Saurabh Sahu, Rahul Gupta, Carol Espy-Wilson

Emotion recognition is a classic field of research with a typical setup extracting features and feeding them through a classifier for prediction. On the other hand, generative models jointly capture the distributional relationship between emotions and the feature profiles. Relatively recently, Generative Adversarial Networks (GANs) have surfaced as a new class of generative models and have shown considerable success in modeling distributions in the fields of computer vision and natural language understanding. In this work, we experiment with variants of GAN architectures to generate feature vectors corresponding to an emotion in two ways: (i) A generator is trained with samples from a mixture prior. Each mixture component corresponds to an emotional class and can be sampled to generate features from the corresponding emotion. (ii) A one-hot vector corresponding to an emotion can be explicitly used to generate the features. We perform analysis on such models and also propose different metrics used to measure the performance of the GAN models in their ability to generate realistic synthetic samples. Apart from evaluation on a given dataset of interest, we perform a cross-corpus study where we study the utility of the synthetic samples as additional training data in low resource conditions.

Saurabh Sahu, Rahul Gupta, Carol Espy-Wilson

Generative Adversarial Networks (GANs) have gained a lot of attention from machine learning community due to their ability to learn and mimic an input data distribution. GANs consist of a discriminator and a generator working in tandem playing a min-max game to learn a target underlying data distribution; when fed with data-points sampled from a simpler distribution (like uniform or Gaussian distribution). Once trained, they allow synthetic generation of examples sampled from the target distribution. We investigate the application of GANs to generate synthetic feature vectors used for speech emotion recognition. Specifically, we investigate two set ups: (i) a vanilla GAN that learns the distribution of a lower dimensional representation of the actual higher dimensional feature vector and, (ii) a conditional GAN that learns the distribution of the higher dimensional feature vectors conditioned on the labels or the emotional class to which it belongs. As a potential practical application of these synthetically generated samples, we measure any improvement in a classifier's performance when the synthetic data is used along with real data for training. We perform cross-validation analyses followed by a cross-corpus study.

Rahul Gupta, Saurabh Sahu, Carol Espy-Wilson et al.

Sentiment classification involves quantifying the affective reaction of a human to a document, media item or an event. Although researchers have investigated several methods to reliably infer sentiment from lexical, speech and body language cues, training a model with a small set of labeled datasets is still a challenge. For instance, in expanding sentiment analysis to new languages and cultures, it may not always be possible to obtain comprehensive labeled datasets. In this paper, we investigate the application of semi-supervised and transfer learning methods to improve performances on low resource sentiment classification tasks. We experiment with extracting dense feature representations, pre-training and manifold regularization in enhancing the performance of sentiment classification systems. Our goal is a coherent implementation of these methods and we evaluate the gains achieved by these methods in matched setting involving training and testing on a single corpus setting as well as two cross corpora settings. In both the cases, our experiments demonstrate that the proposed methods can significantly enhance the model performance against a purely supervised approach, particularly in cases involving a handful of training data.

Saurabh Sahu, Rahul Gupta, Ganesh Sivaraman et al.

Recently, generative adversarial networks and adversarial autoencoders have gained a lot of attention in machine learning community due to their exceptional performance in tasks such as digit classification and face recognition. They map the autoencoder's bottleneck layer output (termed as code vectors) to different noise Probability Distribution Functions (PDFs), that can be further regularized to cluster based on class information. In addition, they also allow a generation of synthetic samples by sampling the code vectors from the mapped PDFs. Inspired by these properties, we investigate the application of adversarial autoencoders to the domain of emotion recognition. Specifically, we conduct experiments on the following two aspects: (i) their ability to encode high dimensional feature vector representations for emotional utterances into a compressed space (with a minimal loss of emotion class discriminability in the compressed space), and (ii) their ability to regenerate synthetic samples in the original feature space, to be later used for purposes such as training emotion recognition classifiers. We demonstrate the promise of adversarial autoencoders with regards to these aspects on the Interactive Emotional Dyadic Motion Capture (IEMOCAP) corpus and present our analysis.