Zhisheng Zheng, Puyuan Peng, Anuj Diwan, Cong Phuoc Huynh, Xiaohang Sun, Zhu Liu, Vimal Bhat, David Harwath. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Target Speech Extraction (TSE) traditionally relies on explicit clues about the speaker’s identity like enrollment audio, face images, or videos, which may not always be available. In this paper, we propose a text-guided TSE model StyleTSE that uses natural language descriptions of speaking style in addition to the audio clue to extract the desired speech from a given mixture. Our model integrates a speech separation network adapted from SepFormer with a bi-modality clue network that flexibly processes both audio and text clues. To train and evaluate our model, we introduce a new dataset TextrolMix with speech mixtures and natural language descriptions. Experimental results demonstrate that our method effectively separates speech based not only on who is speaking, but also on how they are speaking, enhancing TSE when traditional audio clues are absent. Demos are at: https://mingyue66.github.io/TextrolMix/demo/

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.

Long video understanding is inherently challenging for vision-language models (VLMs) because of the extensive number of frames. With each video frame typically expanding into tens or hundreds of tokens, the limited context length of large language models (LLMs) forces the VLMs to perceive the frames sparsely and lose temporal information. To address this, we explore extreme video token compression towards *one token per frame* at the final LLM layer. Our key insight is that heuristic-based compression, widely adopted by previous methods, is prone to information loss, and this necessitates supervising LLM layers into *learnable* and *progressive* modules for *token-level compression* (LP-Comp). Such compression enables our VLM to digest 2x-4x more frames with improved performance. To further increase the token efficiency, we investigate *frame-level compression*, which selects the frames most relevant to the queries via the internal attention scores of the LLM layers, named *question-conditioned compression* (QC-Comp). As a notable distinction from previous studies, we mitigate the position bias of LLM attention in long contexts, *i.e.*, the over-concentration on the beginning and end of a sequence, by splitting long videos into short segments and employing local attention. Collectively, our combined *token-level* and *frame-level* leads to an e**x**treme compression model for long video understanding, named **XComp**, achieving a significantly larger compression ratio and enabling denser frame sampling. Our XComp is finetuned from VideoChat-Flash with a data-efficient *supervised compression tuning* stage that only requires 2.5\% of the supervised fine-tuning data, yet boosts the accuracy from 42.9\% to 46.2\% on LVBench and enhances multiple other long video benchmarks.

Audio Description (AD) plays a pivotal role as an application system aimed at guaranteeing accessibility in multimedia content, which provides additional narrations at suitable intervals to describe visual elements, catering specifically to the needs of visually impaired audiences. In this paper, we introduce CA3D, the pioneering unified ContextAware Automatic Audio Description system that provides AD event scripts with precise locations in the long cinematic content. Specifically, CA3D system consists of: 1) a Temporal Feature Enhancement Module to efficiently capture longer term dependencies, 2) an anchor-based AD event detector with feature suppression module that localizes the AD events and extracts discriminative feature for AD generation, and 3) a self-refinement module that leverages the generated output to tweak AD event boundaries from coarse to fine. Unlike conventional methods which rely on metadata and ground truth AD timestamp for AD detection and generation tasks, the proposed CA3D is the first end-to-end trainable system that only uses visual cue. Extensive experiments demonstrate that the proposed CA3D improves existing architectures for both AD event detection and script generation metrics, establishing the new state-of-the-art performances in the AD automation.

In various video-language learning tasks, the challenge of achieving cross-modality alignment with multi-grained data persists. We propose a method to tackle this challenge from two crucial perspectives: data and modeling. Given the absence of a multi-grained video-text pretraining dataset, we introduce a Granularity EXpansion (GEX) method with Integration and Compression operations to expand the granularity of a single-grained dataset. To better model multi-grained data, we introduce an Iterative Approximation Module (IAM), which embeds multi-grained videos and texts into a unified, low-dimensional semantic space while preserving essential information for cross-modal alignment. Furthermore, GEXIA is highly scalable with no restrictions on the number of video-text granularities for alignment. We evaluate our work on three categories of video tasks across seven benchmark datasets, showcasing state-of-the-art or comparable performance. Remarkably, our model excels in tasks involving long-form video understanding, even though the pretraining dataset only contains short video clips.

Synthesizing high-fidelity talking head videos of an arbitrary identity, lip-synced to a target speech segment, is a challenging problem. Recent GAN-based methods succeed by training a model on a large amount of videos, allowing the generator to learn a variety of audio-lip representations. However, they are unable to handle head pose changes. On the other hand, Neural Radiance Fields (NeRFs) model the 3D face geometry more accurately. Current audio-conditioned NeRFs are not as good in lip synchronization as GANs, since they are trained on limited video data of a single identity. In this work, we propose LipNeRF, a lip-syncing NeRF that bridges the gap between the accurate lip synchronization of GAN-based methods and the accurate 3D face modeling of NeRFs. LipNeRF is conditioned on the expression space of a 3DMM, instead of the audio feature space. We experimentally demonstrate that the expression space gives a better representation for the lip shape than the audio feature space. LipNeRF shows a significant improvement in lip-sync quality over the current state-of-the-art, especially in high-definition videos of cinematic content, with challenging pose, illumination and expression variations.

Several recent works have directly extended the image masked autoencoder (MAE) with random masking into video domain, achieving promising results. However, unlike images, both spatial and temporal information are important for video understanding. This suggests that the random masking strategy that is inherited from the image MAE is less effective for video MAE. This motivates the design of a novel masking algorithm that can more efficiently make use of video saliency. Specifically, we propose a motion-guided masking algorithm (MGM) which leverages motion vectors to guide the position of each mask over time. Crucially, these motion-based correspondences can be directly obtained from information stored in the compressed format of the video, which makes our method efficient and scalable. On two challenging large-scale video benchmarks (Kinetics-400 and Something-Something V2), we equip video MAE with our MGM and achieve up to +1.3% improvement compared to previous state-of-the-art methods. Additionally, our MGM achieves equivalent performance to previous video MAE using up to 66% fewer training epochs. Lastly, we show that MGM generalizes better to downstream transfer learning and domain adaptation tasks on the UCF101, HMDB51, and Diving48 datasets, achieving up to +4.9% improvement compared to baseline methods.

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%).

Scenes play a crucial role in breaking the storyline of movies and TV episodes into semantically cohesive parts. However, given their complex temporal structure, finding scene boundaries can be a challenging task requiring large amounts of labeled training data. To address this challenge, we present a self-supervised shot contrastive learning approach (ShotCoL) to learn a shot representation that maximizes the similarity between nearby shots compared to randomly selected shots. We show how to apply our learned shot representation for the task of scene boundary detection to offer state-of-the-art performance on the MovieNet [33] dataset while requiring only ~25% of the training labels, using 9× fewer model parameters and offering 7× faster runtime. To assess the effectiveness of ShotCoL on novel applications of scene boundary detection, we take on the problem of finding timestamps in movies and TV episodes where video-ads can be inserted while offering a minimally disruptive viewing experience. To this end, we collected a new dataset called AdCuepoints with 3, 975 movies and TV episodes, 2.2 million shots and 19, 119 minimally disruptive ad cue-point labels. We present a thorough empirical analysis on this dataset demonstrating the effectiveness of ShotCoL for ad cue-points detection.