Self-supervised pretraining has transformed speech representation learning, enabling models to generalize across various downstream tasks. However, empirical studies have highlighted two notable gaps. First, different speech tasks require varying levels of acoustic and semantic information, which are encoded at different layers within the model. This adds the extra complexity of layer selection on downstream tasks to reach optimal performance. Second, the entanglement of acoustic and semantic information can undermine model robustness, particularly in varied acoustic environments. To address these issues, we propose a two-branch multitask finetuning strategy that integrates Automatic Speech Recognition and transcript-aligned audio reconstruction, designed to preserve and disentangle semantic and acoustic information in a final layer of a pretrained model. Experiments with the pretrained Wav2Vec 2.0 model demonstrate that our approach surpasses ASR-only finetuning across multiple downstream tasks, and it significantly improves ASR robustness in acoustically varied (emotional) speech.

Danyang Liu, Fanjie Kong, Xiaohang Sun, Dhruva Patil, Avijit Vajpayee, Zhu Liu, Vimal Bhat, Najmeh Sadoughi. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Previous research has studied the task of segmenting cinematic videos into scenes and into narrative acts. However, these studies have overlooked the essential task of multimodal alignment and fusion for effectively and efficiently processing long-form videos (> 60min). In this paper, we introduce Multimodal alignmEnt aGgregation and distillAtion (MEGA) for cinematic long-video segmentation. MEGA tackles the challenge by leveraging multiple media modalities. The method coarsely aligns inputs of variable lengths and different modalities with alignment positional encoding. To maintain temporal synchronization while reducing computation, we further introduce an enhanced bottleneck fusion layer which uses temporal alignment. Additionally, MEGA employs a novel contrastive loss to synchronize and transfer labels across modalities, enabling act segmentation from labeled synopsis sentences on video shots. Our experimental results show that MEGA outperforms state-of-the-art methods on MovieNet dataset for scene segmentation (with an Average Precision improvement of +1.19%) and on TRIPOD dataset for act segmentation (with a Total Agreement improvement of +5.51%).

Articulation, emotion, and personality play strong roles in the orofacial movements. To improve the naturalness and expressiveness ofvirtual agents(VAs), it is important that we carefully model the complex interplay between these factors. This paper proposes a conditional generative adversarial network, calledconditional sequential GAN(CSG), which learns the relationship between emotion, lexical content and lip movements in a principled manner. This model uses a set of spectral and emotional speech features directly extracted from the speech signal as conditioning inputs, generating realistic movements. A key feature of the approach is that it is a speech-driven framework that does not require transcripts. Our experiments show the superiority of this model over three state-of-the-art baselines in terms of objective and subjective evaluations. When the target emotion is known, we propose to create emotionally dependent models by either adapting the base model with the target emotional data (CSG-Emo-Adapted), or adding emotional conditions as the input of the model (CSG-Emo-Aware). Objective evaluations of these models show improvements for the CSG-Emo-Adapted compared with the CSG model, as the trajectory sequences are closer to the original sequences. Subjective evaluations show significantly better results for this model compared with the CSG model when the target emotion is happiness.

The International Classification of Diseases (ICD) is a list of classification codes for the diagnoses. Automatic ICD coding is a multi-label text classification problem with noisy clinical document inputs and long-tailed label distribution, making it difficult for fine-grained classification on both frequent and zero-shot codes at the same time, i.e. generalized zero-shot ICD coding. In this paper, we propose a latent feature generation framework to improve the prediction on unseen codes without compromising the performance on seen codes. Our framework generates semantically meaningful features for zero-shot codes by exploiting ICD code hierarchical structure and reconstructing the code-relevant keywords with a novel cycle architecture. To the best of our knowledge, this is the first adversarial generative model for generalized zero-shot learning on multi-label text classification. Extensive experiments demonstrate the effectiveness of our approach. On the public MIMIC-III dataset, our methods improve the F1 score from nearly 0 to 20.91% for the zero-shot codes, and increase the AUC score by 3% (absolute improvement) from previous state of the art. Code is available at https://github.com/csong27/gzsl_text.

Unsupervised deep hash functions have not shown satisfactory improvements against their shallow alternatives, and usually require supervised pretraining to avoid overfitting. In this paper, we propose a new deep unsupervised hashing function, called HashGAN, which efficiently obtains binary representation of input images without any supervised pretraining. HashGAN consists of three networks, a generator, a discriminator and an encoder. By sharing the parameters of the encoder and discriminator, we benefit from the adversarial loss as a data-dependent regularization in training our deep hash function. Moreover, a novel hashing loss function is introduced for real images, which results in minimum entropy, uniform frequency, consistent and independent hash bits. Furthermore, we employ a collaborative loss in training our model, enforcing similar random inputs and hash bits for synthesized images. In our experiments, HashGAN outperforms the previous unsupervised hash functions in image retrieval and achieves the state-of-the-art performance in image clustering on benchmark datasets. We also provide an ablation study, showing the contribution of each component in our loss function.

The orofacial area conveys a range of information, including speech articulation and emotions. These two factors add constraints to the facial movements, creating non-trivial integrations and interplays. To generate more expressive and naturalistic movements for conversational agents (CAs) the relationship between these factors should be carefully modeled. Data-driven models are more appropriate for this task than rule-based systems. This paper provides two deep learning speech-driven structures to integrate speech articulation and emotional cues. The proposed approaches rely on multitask learning (MTL) strategies, where related secondary tasks are jointly solved when synthesizing orofacial movements. In particular, we evaluate emotion recognition and viseme recognition as secondary tasks. The approach creates shared representations that generate behaviors that not only are closer to the original orofacial movements, but also are perceived more natural than the results from single task learning.

Head movement is an integral part of face-to-face communications. It is important to investigate methodologies to generate naturalistic movements for conversational agents (CAs). The predominant method for head movement generation is using rules based on the meaning of the message. However, the variations of head movements by these methods are bounded by the predefined dictionary of gestures. Speech-driven methods offer an alternative approach, learning the relationship between speech and head movements from real recordings. However, previous studies do not generate novel realizations for a repeated speech signal. Conditional generative adversarial network (GAN) provides a framework to generate multiple realizations of head movements for each speech segment by sampling from a conditioned distribution. We build a conditional GAN with bidirectional long-short term memory (BLSTM), which is suitable for capturing the long-short term dependencies of time-continuous signals. This model learns the distribution of head movements conditioned on speech prosodic features. We compare this model with a dynamic Bayesian network (DBN) and BLSTM models optimized to reduce mean squared error (MSE) or to increase concordance correlation. The objective evaluations and subjective evaluations of the results showed better performance for the conditional GAN model compared with these baseline systems.

Synchrony is an essential aspect of human-human interactions. In previous work, we have seen how synchrony manifests in low-level acoustic phenomena like fundamental frequency, loudness, and the duration of keywords during the play of child-child pairs in a fast-paced, cooperative, language-based game. The correlation between the increase in such low-level synchrony and increase in enjoyment of the game suggests that a similar dynamic between child and robot co-players might also improve the child's experience. We report an approach to creating on-line acoustic synchrony by using a dynamic Bayesian network learned from prior recordings of child-child play to select from a predefined space of robot speech in response to real-time measurement of the child's prosodic features. Data were collected from 40 new children, each playing the game with both a synchronizing and non-synchronizing version of the robot. Results show a significant order effect: although all children grew to enjoy the game more over time, those that began with the synchronous robot maintained their own synchrony to it and achieved higher engagement compared with those that did not.

We present the MSP-IMPROV corpus, a multimodal emotional database, where the goal is to have control over lexical content and emotion while also promoting naturalness in the recordings. Studies on emotion perception often require stimuli with fixed lexical content, but that convey different emotions. These stimuli can also serve as an instrument to understand how emotion modulates speech at the phoneme level, in a manner that controls for coarticulation. Such audiovisual data are not easily available from natural recordings. A common solution is to record actors reading sentences that portray different emotions, which may not produce natural behaviors. We propose an alternative approach in which we define hypothetical scenarios for each sentence that are carefully designed to elicit a particular emotion. Two actors improvise these emotion-specific situations, leading them to utter contextualized, non-read renditions of sentences that have fixed lexical content and convey different emotions. We describe the context in which this corpus was recorded, the key features of the corpus, the areas in which this corpus can be useful, and the emotional content of the recordings. The paper also provides the performance for speech and facial emotion classifiers. The analysis brings novel classification evaluations where we study the performance in terms of inter-evaluator agreement and naturalness perception, leveraging the large size of the audiovisual database.

Creating believable behaviors for conversational agents (CAs) is a challenging task, given the complex relationship between speech and various nonverbal behaviors. The two main approaches are rule-based systems, which tend to produce behaviors with limited variations compared to natural interactions, and data-driven systems, which tend to ignore the underlying semantic meaning of the message (e.g., gestures without meaning). We envision a hybrid system, acting as the behavior realization layer in rule-based systems, while exploiting the rich variation in natural interactions. Constrained on a given target gesture (e.g., head nod) and speech signal, the system will generate novel realizations learned from the data, capturing the timely relationship between speech and gestures. An important task in this research is identifying multiple examples of the target gestures in the corpus. This paper proposes a data mining framework for detecting gestures of interest in a motion capture database. First, we train One-class support vector machines (SVMs) to detect candidate segments conveying the target gesture. Second, we use dynamic time alignment kernel (DTAK) to compare the similarity between the examples (i.e., target gesture) and the given segments. We evaluate the approach for five prototypical hand and head gestures showing reasonable performance. These retrieved gestures are then used to train a speech-driven framework based on dynamic Bayesian networks (DBNs) to synthesize these target behaviors.

Conversational agents provide powerful opportunities to interact and engage with the users. The challenge is how to create naturalistic behaviors that replicate the complex gestures observed during human interactions. Previous studies have used rule-based frameworks or data-driven models to generate appropriate gestures, which are properly synchronized with the underlying discourse functions. Among these methods, speech-driven approaches are especially appealing given the rich information conveyed on speech. It captures emotional cues and prosodic patterns that are important to synthesize behaviors (i.e., modeling the variability and complexity of the timings of the behaviors). The main limitation of these models is that they fail to capture the underlying semantic and discourse functions of the message (e.g., nodding). This study proposes a speech-driven framework that explicitly model discourse functions, bridging the gap between speech-driven and rule-based models. The approach is based on dynamic Bayesian Network (DBN), where an additional node is introduced to constrain the models by specific discourse functions. We implement the approach by synthesizing head and eyebrow motion. We conduct perceptual evaluations to compare the animations generated using the constrained and unconstrained models.