Bowen Shi

Mohamed Anwar, Bowen Shi, Vedanuj Goswami et al. · meta-ai

We introduce MuAViC, a multilingual audio-visual corpus for robust speech recognition and robust speech-to-text translation providing 1200 hours of audio-visual speech in 9 languages. It is fully transcribed and covers 6 English-to-X translation as well as 6 X-to-English translation directions. To the best of our knowledge, this is the first open benchmark for audio-visual speech-to-text translation and the largest open benchmark for multilingual audio-visual speech recognition. Our baseline results show that MuAViC is effective for building noise-robust speech recognition and translation models. We make the corpus available at https://github.com/facebookresearch/muavic.

Lili Yu, Bowen Shi, Ramakanth Pasunuru et al. · berkeley, meta-ai

We present CM3Leon (pronounced "Chameleon"), a retrieval-augmented, token-based, decoder-only multi-modal language model capable of generating and infilling both text and images. CM3Leon uses the CM3 multi-modal architecture but additionally shows the extreme benefits of scaling up and tuning on more diverse instruction-style data. It is the first multi-modal model trained with a recipe adapted from text-only language models, including a large-scale retrieval-augmented pre-training stage and a second multi-task supervised fine-tuning (SFT) stage. It is also a general-purpose model that can do both text-to-image and image-to-text generation, allowing us to introduce self-contained contrastive decoding methods that produce high-quality outputs. Extensive experiments demonstrate that this recipe is highly effective for multi-modal models. CM3Leon achieves state-of-the-art performance in text-to-image generation with 5x less training compute than comparable methods (zero-shot MS-COCO FID of 4.88). After SFT, CM3Leon can also demonstrate unprecedented levels of controllability in tasks ranging from language-guided image editing to image-controlled generation and segmentation.

Tu Anh Nguyen, Wei-Ning Hsu, Antony D'Avirro et al.

Recent work has shown that it is possible to resynthesize high-quality speech based, not on text, but on low bitrate discrete units that have been learned in a self-supervised fashion and can therefore capture expressive aspects of speech that are hard to transcribe (prosody, voice styles, non-verbal vocalization). The adoption of these methods is still limited by the fact that most speech synthesis datasets are read, severely limiting spontaneity and expressivity. Here, we introduce Expresso, a high-quality expressive speech dataset for textless speech synthesis that includes both read speech and improvised dialogues rendered in 26 spontaneous expressive styles. We illustrate the challenges and potentials of this dataset with an expressive resynthesis benchmark where the task is to encode the input in low-bitrate units and resynthesize it in a target voice while preserving content and style. We evaluate resynthesis quality with automatic metrics for different self-supervised discrete encoders, and explore tradeoffs between quality, bitrate and invariance to speaker and style. All the dataset, evaluation metrics and baseline models are open source

Bowen Shi, Diane Brentari, Greg Shakhnarovich et al.

Existing work on sign language translation - that is, translation from sign language videos into sentences in a written language - has focused mainly on (1) data collected in a controlled environment or (2) data in a specific domain, which limits the applicability to real-world settings. In this paper, we introduce OpenASL, a large-scale American Sign Language (ASL) - English dataset collected from online video sites (e.g., YouTube). OpenASL contains 288 hours of ASL videos in multiple domains from over 200 signers and is the largest publicly available ASL translation dataset to date. To tackle the challenges of sign language translation in realistic settings and without glosses, we propose a set of techniques including sign search as a pretext task for pre-training and fusion of mouthing and handshape features. The proposed techniques produce consistent and large improvements in translation quality, over baseline models based on prior work. Our data and code are publicly available at https://github.com/chevalierNoir/OpenASL

Wei-Ning Hsu, Bowen Shi

While audio-visual speech models can yield superior performance and robustness compared to audio-only models, their development and adoption are hindered by the lack of labeled and unlabeled audio-visual data and the cost to deploy one model per modality. In this paper, we present u-HuBERT, a self-supervised pre-training framework that can leverage both multimodal and unimodal speech with a unified masked cluster prediction objective. By utilizing modality dropout during pre-training, we demonstrate that a single fine-tuned model can achieve performance on par or better than the state-of-the-art modality-specific models. Moreover, our model fine-tuned only on audio can perform well with audio-visual and visual speech input, achieving zero-shot modality generalization for multiple speech processing tasks. In particular, our single model yields 1.2%/1.4%/27.2% speech recognition word error rate on LRS3 with audio-visual/audio/visual input. Codes and models are available at https://github.com/facebookresearch/av_hubert

Jin Li, Yaoming Wang, Xiaopeng Zhang et al.

Vision transformers (ViTs) have emerged as a prevalent architecture for vision tasks owing to their impressive performance. However, when it comes to handling long token sequences, especially in dense prediction tasks that require high-resolution input, the complexity of ViTs increases significantly. Notably, dense prediction tasks, such as semantic segmentation or object detection, emphasize more on the contours or shapes of objects, while the texture inside objects is less informative. Motivated by this observation, we propose to apply adaptive resolution for different regions in the image according to their importance. Specifically, at the intermediate layer of the ViT, we utilize a spatial-aware density-based clustering algorithm to select representative tokens from the token sequence. Once the representative tokens are determined, we proceed to merge other tokens into their closest representative token. Consequently, semantic similar tokens are merged together to form low-resolution regions, while semantic irrelevant tokens are preserved independently as high-resolution regions. This strategy effectively reduces the number of tokens, allowing subsequent layers to handle a reduced token sequence and achieve acceleration. We evaluate our proposed method on three different datasets and observe promising performance. For example, the "Segmenter ViT-L" model can be accelerated by 48% FPS without fine-tuning, while maintaining the performance. Additionally, our method can be applied to accelerate fine-tuning as well. Experimental results demonstrate that we can save 52% training time while accelerating 2.46 times FPS with only a 0.09% performance drop. The code is available at https://github.com/caddyless/ailurus/tree/main.

Matthew Le, Apoorv Vyas, Bowen Shi et al.

Large-scale generative models such as GPT and DALL-E have revolutionized the research community. These models not only generate high fidelity outputs, but are also generalists which can solve tasks not explicitly taught. In contrast, speech generative models are still primitive in terms of scale and task generalization. In this paper, we present Voicebox, the most versatile text-guided generative model for speech at scale. Voicebox is a non-autoregressive flow-matching model trained to infill speech, given audio context and text, trained on over 50K hours of speech that are not filtered or enhanced. Similar to GPT, Voicebox can perform many different tasks through in-context learning, but is more flexible as it can also condition on future context. Voicebox can be used for mono or cross-lingual zero-shot text-to-speech synthesis, noise removal, content editing, style conversion, and diverse sample generation. In particular, Voicebox outperforms the state-of-the-art zero-shot TTS model VALL-E on both intelligibility (5.9% vs 1.9% word error rates) and audio similarity (0.580 vs 0.681) while being up to 20 times faster. Audio samples can be found in \url{https://voicebox.metademolab.com}.

Ankita Pasad, Bowen Shi, Karen Livescu

Many self-supervised speech models, varying in their pre-training objective, input modality, and pre-training data, have been proposed in the last few years. Despite impressive successes on downstream tasks, we still have a limited understanding of the properties encoded by the models and the differences across models. In this work, we examine the intermediate representations for a variety of recent models. Specifically, we measure acoustic, phonetic, and word-level properties encoded in individual layers, using a lightweight analysis tool based on canonical correlation analysis (CCA). We find that these properties evolve across layers differently depending on the model, and the variations relate to the choice of pre-training objective. We further investigate the utility of our analyses for downstream tasks by comparing the property trends with performance on speech recognition and spoken language understanding tasks. We discover that CCA trends provide reliable guidance to choose layers of interest for downstream tasks and that single-layer performance often matches or improves upon using all layers, suggesting implications for more efficient use of pre-trained models.

Helin Wang, Bowen Shi, Andros Tjandra et al.

The performance evaluation remains a complex challenge in audio separation, and existing evaluation metrics are often misaligned with human perception, course-grained, relying on ground truth signals. On the other hand, subjective listening tests remain the gold standard for real-world evaluation, but they are expensive, time-consuming, and difficult to scale. This paper addresses the growing need for automated systems capable of evaluating audio separation without human intervention. The proposed evaluation metric, SAM Audio Judge (SAJ), is a multimodal fine-grained reference-free objective metric, which shows highly alignment with human perceptions. SAJ supports three audio domains (speech, music and general sound events) and three prompt inputs (text, visual and span), covering four different dimensions of evaluation (recall, percision, faithfulness, and overall). SAM Audio Judge also shows potential applications in data filtering, pseudo-labeling large datasets and reranking in audio separation models. We release our code and pre-trained models at: https://github.com/facebookresearch/sam-audio.

Alexander H. Liu, Matt Le, Apoorv Vyas et al.

Generative models have gained more and more attention in recent years for their remarkable success in tasks that required estimating and sampling data distribution to generate high-fidelity synthetic data. In speech, text-to-speech synthesis and neural vocoder are good examples where generative models have shined. While generative models have been applied to different applications in speech, there exists no general-purpose generative model that models speech directly. In this work, we take a step toward this direction by showing a single pre-trained generative model can be adapted to different downstream tasks with strong performance. Specifically, we pre-trained a generative model, named SpeechFlow, on 60k hours of untranscribed speech with Flow Matching and masked conditions. Experiment results show the pre-trained generative model can be fine-tuned with task-specific data to match or surpass existing expert models on speech enhancement, separation, and synthesis. Our work suggested a foundational model for generation tasks in speech can be built with generative pre-training.

Han Li, Bowen Shi, Wenrui Dai et al.

There has been a recent surge of interest in introducing transformers to 3D human pose estimation (HPE) due to their powerful capabilities in modeling long-term dependencies. However, existing transformer-based methods treat body joints as equally important inputs and ignore the prior knowledge of human skeleton topology in the self-attention mechanism. To tackle this issue, in this paper, we propose a Pose-Oriented Transformer (POT) with uncertainty guided refinement for 3D HPE. Specifically, we first develop novel pose-oriented self-attention mechanism and distance-related position embedding for POT to explicitly exploit the human skeleton topology. The pose-oriented self-attention mechanism explicitly models the topological interactions between body joints, whereas the distance-related position embedding encodes the distance of joints to the root joint to distinguish groups of joints with different difficulties in regression. Furthermore, we present an Uncertainty-Guided Refinement Network (UGRN) to refine pose predictions from POT, especially for the difficult joints, by considering the estimated uncertainty of each joint with uncertainty-guided sampling strategy and self-attention mechanism. Extensive experiments demonstrate that our method significantly outperforms the state-of-the-art methods with reduced model parameters on 3D HPE benchmarks such as Human3.6M and MPI-INF-3DHP

Ning Liao, Bowen Shi, Xiaopeng Zhang et al.

Prompt learning has achieved great success in efficiently exploiting large-scale pre-trained models in natural language processing (NLP). It reformulates the downstream tasks as the generative pre-training ones to achieve consistency, thus improving the performance stably. However, when transferring it to the vision area, current visual prompt learning methods are almost designed on discriminative pre-trained models, and there is also a lack of careful design to unify the forms of pre-training and downstream tasks. To explore prompt learning on the generative pre-trained visual model, as well as keeping the task consistency, we propose Visual Prompt learning as masked visual Token Modeling (VPTM) to transform the downstream visual classification into the pre-trained masked visual token prediction. In addition, we develop the prototypical verbalizer for mapping the predicted visual token with implicit semantics to explicit downstream labels. To our best knowledge, VPTM is the first visual prompt method on the generative pre-trained visual model, which achieves consistency between pre-training and downstream visual classification by task reformulation. Experiments show that VPTM outperforms other visual prompt methods and achieves excellent efficiency. Moreover, the task consistency of VPTM contributes to the robustness against prompt location, prompt length and prototype dimension, and could be deployed uniformly.

Bowen Shi, Abdelrahman Mohamed, Wei-Ning Hsu

This paper investigates self-supervised pre-training for audio-visual speaker representation learning where a visual stream showing the speaker's mouth area is used alongside speech as inputs. Our study focuses on the Audio-Visual Hidden Unit BERT (AV-HuBERT) approach, a recently developed general-purpose audio-visual speech pre-training framework. We conducted extensive experiments probing the effectiveness of pre-training and visual modality. Experimental results suggest that AV-HuBERT generalizes decently to speaker related downstream tasks, improving label efficiency by roughly ten fold for both audio-only and audio-visual speaker verification. We also show that incorporating visual information, even just the lip area, greatly improves the performance and noise robustness, reducing EER by 38% in the clean condition and 75% in noisy conditions.

Bowen Shi, Diane Brentari, Greg Shakhnarovich et al.

Natural language processing for sign language video - including tasks like recognition, translation, and search - is crucial for making artificial intelligence technologies accessible to deaf individuals, and is gaining research interest in recent years. In this paper, we address the problem of searching for fingerspelled key-words or key phrases in raw sign language videos. This is an important task since significant content in sign language is often conveyed via fingerspelling, and to our knowledge the task has not been studied before. We propose an end-to-end model for this task, FSS-Net, that jointly detects fingerspelling and matches it to a text sequence. Our experiments, done on a large public dataset of ASL fingerspelling in the wild, show the importance of fingerspelling detection as a component of a search and retrieval model. Our model significantly outperforms baseline methods adapted from prior work on related tasks

Wei-Ning Hsu, Tal Remez, Bowen Shi et al.

Prior works on improving speech quality with visual input typically study each type of auditory distortion separately (e.g., separation, inpainting, video-to-speech) and present tailored algorithms. This paper proposes to unify these subjects and study Generalized Speech Enhancement, where the goal is not to reconstruct the exact reference clean signal, but to focus on improving certain aspects of speech. In particular, this paper concerns intelligibility, quality, and video synchronization. We cast the problem as audio-visual speech resynthesis, which is composed of two steps: pseudo audio-visual speech recognition (P-AVSR) and pseudo text-to-speech synthesis (P-TTS). P-AVSR and P-TTS are connected by discrete units derived from a self-supervised speech model. Moreover, we utilize self-supervised audio-visual speech model to initialize P-AVSR. The proposed model is coined ReVISE. ReVISE is the first high-quality model for in-the-wild video-to-speech synthesis and achieves superior performance on all LRS3 audio-visual enhancement tasks with a single model. To demonstrates its applicability in the real world, ReVISE is also evaluated on EasyCom, an audio-visual benchmark collected under challenging acoustic conditions with only 1.6 hours of training data. Similarly, ReVISE greatly suppresses noise and improves quality. Project page: https://wnhsu.github.io/ReVISE.

Hongwei Zheng, Han Li, Bowen Shi et al.

Recent 2D-to-3D human pose estimation (HPE) utilizes temporal consistency across sequences to alleviate the depth ambiguity problem but ignore the action related prior knowledge hidden in the pose sequence. In this paper, we propose a plug-and-play module named Action Prompt Module (APM) that effectively mines different kinds of action clues for 3D HPE. The highlight is that, the mining scheme of APM can be widely adapted to different frameworks and bring consistent benefits. Specifically, we first present a novel Action-related Text Prompt module (ATP) that directly embeds action labels and transfers the rich language information in the label to the pose sequence. Besides, we further introduce Action-specific Pose Prompt module (APP) to mine the position-aware pose pattern of each action, and exploit the correlation between the mined patterns and input pose sequence for further pose refinement. Experiments show that APM can improve the performance of most video-based 2D-to-3D HPE frameworks by a large margin.

Bowen Shi, Xiaopeng Zhang, Yaoming Wang et al.

Representation learning has been evolving from traditional supervised training to Contrastive Learning (CL) and Masked Image Modeling (MIM). Previous works have demonstrated their pros and cons in specific scenarios, i.e., CL and supervised pre-training excel at capturing longer-range global patterns and enabling better feature discrimination, while MIM can introduce more local and diverse attention across all transformer layers. In this paper, we explore how to obtain a model that combines their strengths. We start by examining previous feature distillation and mask feature reconstruction methods and identify their limitations. We find that their increasing diversity mainly derives from the asymmetric designs, but these designs may in turn compromise the discrimination ability. In order to better obtain both discrimination and diversity, we propose a simple but effective Hybrid Distillation strategy, which utilizes both the supervised/CL teacher and the MIM teacher to jointly guide the student model. Hybrid Distill imitates the token relations of the MIM teacher to alleviate attention collapse, as well as distills the feature maps of the supervised/CL teacher to enable discrimination. Furthermore, a progressive redundant token masking strategy is also utilized to reduce the distilling costs and avoid falling into local optima. Experiment results prove that Hybrid Distill can achieve superior performance on different benchmarks.

Bowen Shi, Andros Tjandra, John Hoffman et al.

General audio source separation is a key capability for multimodal AI systems that can perceive and reason about sound. Despite substantial progress in recent years, existing separation models are either domain-specific, designed for fixed categories such as speech or music, or limited in controllability, supporting only a single prompting modality such as text. In this work, we present SAM Audio, a foundation model for general audio separation that unifies text, visual, and temporal span prompting within a single framework. Built on a diffusion transformer architecture, SAM Audio is trained with flow matching on large-scale audio data spanning speech, music, and general sounds, and can flexibly separate target sources described by language, visual masks, or temporal spans. The model achieves state-of-the-art performance across a diverse suite of benchmarks, including general sound, speech, music, and musical instrument separation in both in-the-wild and professionally produced audios, substantially outperforming prior general-purpose and specialized systems. Furthermore, we introduce a new real-world separation benchmark with human-labeled multimodal prompts and a reference-free evaluation model that correlates strongly with human judgment.

Leying Zhang, Bowen Shi, Haibin Wu et al.

The rapid advancement of generative audio models has outpaced the development of robust evaluation methodologies. Existing objective metrics and general multimodal large language models (MLLMs) often struggle with domain generalization, zero-shot capabilities, and instructional flexibility. To address these bottlenecks, we propose JASTIN, a generalizable, instruction-driven audio evaluation framework that formulates audio assessment as a self-instructed reasoning task. JASTIN bridges a frozen high-performance audio encoder with a fine-tuned LLM backbone via a trainable audio adapter. To ensure robust zero-shot generalization, we introduce a comprehensive instruction following data preparation pipeline, incorporating Multi-Source, Multi-Task, Multi-Calibration, and Multi-Description data. Experimental results demonstrate that JASTIN achieves state-of-the-art Pearson and Spearman correlations with human subjective ratings. It consistently outperforms general MLLMs across speech, sound, music, and out-of-domain evaluation tasks without the need for task-specific retraining.

Yuetian Chen, Bowen Shi, Mei Si

Artificial intelligence has made significant progress in natural language processing, with models like GPT-3 demonstrating impressive capabilities. However, these models still have limitations when it comes to complex tasks that require an understanding of the user, such as mastering human comedy writing strategies. This paper explores humor generation using GPT-3 by modeling human comedy writing theory and leveraging step-by-step thinking instructions. In addition, we explore the role of cognitive distance in creating humor.

Andros Tjandra, Yi-Chiao Wu, Baishan Guo et al.

The quantification of audio aesthetics remains a complex challenge in audio processing, primarily due to its subjective nature, which is influenced by human perception and cultural context. Traditional methods often depend on human listeners for evaluation, leading to inconsistencies and high resource demands. This paper addresses the growing need for automated systems capable of predicting audio aesthetics without human intervention. Such systems are crucial for applications like data filtering, pseudo-labeling large datasets, and evaluating generative audio models, especially as these models become more sophisticated. In this work, we introduce a novel approach to audio aesthetic evaluation by proposing new annotation guidelines that decompose human listening perspectives into four distinct axes. We develop and train no-reference, per-item prediction models that offer a more nuanced assessment of audio quality. Our models are evaluated against human mean opinion scores (MOS) and existing methods, demonstrating comparable or superior performance. This research not only advances the field of audio aesthetics but also provides open-source models and datasets to facilitate future work and benchmarking. We release our code and pre-trained model at: https://github.com/facebookresearch/audiobox-aesthetics

Yuetian Chen, Ruohua Li, Bowen Shi et al.

Automated visual story generation aims to produce stories with corresponding illustrations that exhibit coherence, progression, and adherence to characters' emotional development. This work proposes a story generation pipeline to co-create visual stories with the users. The pipeline allows the user to control events and emotions on the generated content. The pipeline includes two parts: narrative and image generation. For narrative generation, the system generates the next sentence using user-specified keywords and emotion labels. For image generation, diffusion models are used to create a visually appealing image corresponding to each generated sentence. Further, object recognition is applied to the generated images to allow objects in these images to be mentioned in future story development.

Bowen Shi

Sign language, which conveys meaning through gestures, is the chief means of communication among deaf people. Recognizing sign language in natural settings presents significant challenges due to factors such as lighting, background clutter, and variations in signer characteristics. In this thesis, I study automatic sign language processing in the wild, using signing videos collected from the Internet. This thesis contributes new datasets, tasks, and methods. Most chapters of this thesis address tasks related to fingerspelling, an important component of sign language and yet has not been studied widely by prior work. I present three new large-scale ASL datasets in the wild: ChicagoFSWild, ChicagoFSWild+, and OpenASL. Using ChicagoFSWild and ChicagoFSWild+, I address fingerspelling recognition, which consists of transcribing fingerspelling sequences into text. I propose an end-to-end approach based on iterative attention that allows recognition from a raw video without explicit hand detection. I further show that using a Conformer-based network jointly modeling handshape and mouthing can bring performance close to that of humans. Next, I propose two tasks for building real-world fingerspelling-based applications: fingerspelling detection and search. For fingerspelling detection, I introduce a suite of evaluation metrics and a new detection model via multi-task training. To address the problem of searching for fingerspelled keywords in raw sign language videos, we propose a novel method that jointly localizes and matches fingerspelling segments to text. Finally, I will describe a benchmark for large-vocabulary open-domain sign language translation based on OpenASL. To address the challenges of sign language translation in realistic settings, we propose a set of techniques including sign search as a pretext task for pre-training and fusion of mouthing and handshape features.

Yue Hou, Xueyuan Chen, He Zhu et al.

Novel Class Discovery (NCD) involves identifying new categories within unlabeled data by utilizing knowledge acquired from previously established categories. However, existing NCD methods often struggle to maintain a balance between the performance of old and new categories. Discovering unlabeled new categories in a class-incremental way is more practical but also more challenging, as it is frequently hindered by either catastrophic forgetting of old categories or an inability to learn new ones. Furthermore, the implementation of NCD on continuously scalable graph-structured data remains an under-explored area. In response to these challenges, we introduce for the first time a more practical NCD scenario for node classification (i.e., NC-NCD), and propose a novel self-training framework with prototype replay and distillation called SWORD, adopted to our NC-NCD setting. Our approach enables the model to cluster unlabeled new category nodes after learning labeled nodes while preserving performance on old categories without reliance on old category nodes. SWORD achieves this by employing a self-training strategy to learn new categories and preventing the forgetting of old categories through the joint use of feature prototypes and knowledge distillation. Extensive experiments on four common benchmarks demonstrate the superiority of SWORD over other state-of-the-art methods.

Bowen Shi, Peisen Zhao, Zichen Wang et al.

Vision-language foundation models, represented by Contrastive Language-Image Pre-training (CLIP), have gained increasing attention for jointly understanding both vision and textual tasks. However, existing approaches primarily focus on training models to match global image representations with textual descriptions, thereby overlooking the critical alignment between local regions and corresponding text tokens. This paper extends CLIP with multi-granularity alignment. Notably, we deliberately construct a new dataset comprising pseudo annotations at various levels of granularities, encompassing image-level, region-level as well as pixel-level captions and tags. Accordingly, we develop a Unified Multi-Granularity learning framework, termed UMG-CLIP, which simultaneously empowers the model with versatile perception abilities across different levels of detail. With parameter efficient tuning, UMG-CLIP surpasses current widely used CLIP variants and achieves state-of-the-art performance on diverse image understanding benchmarks, including open-world recognition, retrieval, semantic segmentation, and panoptic segmentation tasks. We believe that UMG-CLIP represents a valuable advancement in vision-language foundation models. The code is available at https://github.com/lygsbw/UMG-CLIP.

Gabrielle Kaili-May Liu, Bowen Shi, Avi Caciularu et al.

Multi-document (MD) processing is crucial for LLMs to handle real-world tasks such as summarization and question-answering across large sets of documents. While LLMs have improved at processing long inputs, MD contexts still present unique difficulties, including management of inter-document dependencies, redundancy, and incoherent structures. To address this challenge, we introduce MDCure, a scalable and effective instruction data generation framework to enhance the MD capabilities of LLMs without the computational cost of pre-training or reliance on human-annotated data. MDCure generates high-quality synthetic MD instruction data over sets of articles via targeted prompts. We also introduce MDCureRM, a cost-effective, MD-specific reward model to score and filter generated data based on their training utility for MD settings. MDCure is compatible with open- and closed-source models in addition to policy optimization methods such as PPO, enabling even small open-source models to surpass proprietary LLMs as strong generators of high-quality MD instruction data without further data filtering. With MDCure, we fine-tune a wide variety of LLMs up to 70B parameters in size from the FlanT5, Qwen2, and LLAMA3.1 model families. Extensive evaluations on a wide range of MD and long-context benchmarks spanning various tasks and domains show MDCure consistently improves performance over pre-trained baselines and base models by up to 75.1%. Our code, datasets, and models are available at https://github.com/yale-nlp/MDCure.

K R Prajwal, Bowen Shi, Matthew Lee et al. · oxford

We introduce MusicFlow, a cascaded text-to-music generation model based on flow matching. Based on self-supervised representations to bridge between text descriptions and music audios, we construct two flow matching networks to model the conditional distribution of semantic and acoustic features. Additionally, we leverage masked prediction as the training objective, enabling the model to generalize to other tasks such as music infilling and continuation in a zero-shot manner. Experiments on MusicCaps reveal that the music generated by MusicFlow exhibits superior quality and text coherence despite being over $2\sim5$ times smaller and requiring $5$ times fewer iterative steps. Simultaneously, the model can perform other music generation tasks and achieves competitive performance in music infilling and continuation. Our code and model will be publicly available.

Adam Polyak, Amit Zohar, Andrew Brown et al. · meta-ai

We present Movie Gen, a cast of foundation models that generates high-quality, 1080p HD videos with different aspect ratios and synchronized audio. We also show additional capabilities such as precise instruction-based video editing and generation of personalized videos based on a user's image. Our models set a new state-of-the-art on multiple tasks: text-to-video synthesis, video personalization, video editing, video-to-audio generation, and text-to-audio generation. Our largest video generation model is a 30B parameter transformer trained with a maximum context length of 73K video tokens, corresponding to a generated video of 16 seconds at 16 frames-per-second. We show multiple technical innovations and simplifications on the architecture, latent spaces, training objectives and recipes, data curation, evaluation protocols, parallelization techniques, and inference optimizations that allow us to reap the benefits of scaling pre-training data, model size, and training compute for training large scale media generation models. We hope this paper helps the research community to accelerate progress and innovation in media generation models. All videos from this paper are available at https://go.fb.me/MovieGenResearchVideos.

Yuchen Liu, Yaoming Wang, Bowen Shi et al.

Vision encoders serve as the cornerstone of multimodal understanding. Single-encoder architectures like CLIP exhibit inherent constraints in generalizing across diverse multimodal tasks, while recent multi-encoder fusion methods introduce prohibitive computational overhead to achieve superior performance using complementary visual representations from multiple vision encoders. To address this, we propose a progressive pruning framework, namely Multi-Encoder collaboraTivE tOken pRuning (METEOR), that eliminates redundant visual tokens across the encoding, fusion, and decoding stages for multi-encoder MLLMs. For multi-vision encoding, we discard redundant tokens within each encoder via a rank guided collaborative token assignment strategy. Subsequently, for multi-vision fusion, we combine the visual features from different encoders while reducing cross-encoder redundancy with cooperative pruning. Finally, we propose an adaptive token pruning method in the LLM decoding stage to further discard irrelevant tokens based on the text prompts with dynamically adjusting pruning ratios for specific task demands. To our best knowledge, this is the first successful attempt that achieves an efficient multi-encoder based vision language model with multi-stage pruning strategies. Extensive experiments on 11 benchmarks demonstrate the effectiveness of our proposed approach. Compared with EAGLE, a typical multi-encoder MLLMs, METEOR reduces 76% visual tokens with only 0.3% performance drop in average. The code is available at https://github.com/YuchenLiu98/METEOR.

Kaibo Hu, Ting Lin, Bowen Shi

We construct a family of finite element sub-complexes of the conformal complex on tetrahedral meshes and show their exactness on contractible domains. This complex includes vector fields and symmetric and traceless tensor fields, connected through the conformal Killing operator, the linearized Cotton-York operator, and the divergence operator, respectively. This leads to discrete versions of transverse traceless (TT) tensors, i.e., symmetric, traceless and divergence-free matrix fields, in continuum mechanics and general relativity. We also show the inf-sup stability of the $H(\operatorname{div})$-conforming finite element symmetric and traceless tensors paired with discontinuous vectors.

Apoorv Vyas, Bowen Shi, Matthew Le et al.

Audio is an essential part of our life, but creating it often requires expertise and is time-consuming. Research communities have made great progress over the past year advancing the performance of large scale audio generative models for a single modality (speech, sound, or music) through adopting more powerful generative models and scaling data. However, these models lack controllability in several aspects: speech generation models cannot synthesize novel styles based on text description and are limited on domain coverage such as outdoor environments; sound generation models only provide coarse-grained control based on descriptions like "a person speaking" and would only generate mumbling human voices. This paper presents Audiobox, a unified model based on flow-matching that is capable of generating various audio modalities. We design description-based and example-based prompting to enhance controllability and unify speech and sound generation paradigms. We allow transcript, vocal, and other audio styles to be controlled independently when generating speech. To improve model generalization with limited labels, we adapt a self-supervised infilling objective to pre-train on large quantities of unlabeled audio. Audiobox sets new benchmarks on speech and sound generation (0.745 similarity on Librispeech for zero-shot TTS; 0.77 FAD on AudioCaps for text-to-sound) and unlocks new methods for generating audio with novel vocal and acoustic styles. We further integrate Bespoke Solvers, which speeds up generation by over 25 times compared to the default ODE solver for flow-matching, without loss of performance on several tasks. Our demo is available at https://audiobox.metademolab.com/

Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia et al.

Video recordings of speech contain correlated audio and visual information, providing a strong signal for speech representation learning from the speaker's lip movements and the produced sound. We introduce Audio-Visual Hidden Unit BERT (AV-HuBERT), a self-supervised representation learning framework for audio-visual speech, which masks multi-stream video input and predicts automatically discovered and iteratively refined multimodal hidden units. AV-HuBERT learns powerful audio-visual speech representation benefiting both lip-reading and automatic speech recognition. On the largest public lip-reading benchmark LRS3 (433 hours), AV-HuBERT achieves 32.5% WER with only 30 hours of labeled data, outperforming the former state-of-the-art approach (33.6%) trained with a thousand times more transcribed video data (31K hours). The lip-reading WER is further reduced to 26.9% when using all 433 hours of labeled data from LRS3 and combined with self-training. Using our audio-visual representation on the same benchmark for audio-only speech recognition leads to a 40% relative WER reduction over the state-of-the-art performance (1.3% vs 2.3%). Our code and models are available at https://github.com/facebookresearch/av_hubert

Han Li, Bowen Shi, Wenrui Dai et al.

Recent 2D-to-3D human pose estimation works tend to utilize the graph structure formed by the topology of the human skeleton. However, we argue that this skeletal topology is too sparse to reflect the body structure and suffer from serious 2D-to-3D ambiguity problem. To overcome these weaknesses, we propose a novel graph convolution network architecture, Hierarchical Graph Networks (HGN). It is based on denser graph topology generated by our multi-scale graph structure building strategy, thus providing more delicate geometric information. The proposed architecture contains three sparse-to-fine representation subnetworks organized in parallel, in which multi-scale graph-structured features are processed and exchange information through a novel feature fusion strategy, leading to rich hierarchical representations. We also introduce a 3D coarse mesh constraint to further boost detail-related feature learning. Extensive experiments demonstrate that our HGN achieves the state-of-the art performance with reduced network parameters. Code is released at https://github.com/qingshi9974/BMVC2021-Hierarchical-Graph-Networks-for-3D-Human-Pose-Estimation.

Ying Wang, Bihuan Chen, Kaifeng Huang et al.

Third-party libraries are a central building block to develop software systems. However, outdated third-party libraries are commonly used, and developers are usually less aware of the potential risks. Therefore, a quantitative and holistic study on usages, updates and risks of third-party libraries can provide practical insights to improve the ecosystem sustainably. In this paper, we conduct such a study in the Java ecosystem. Specifically, we conduct a library usage analysis (e.g., usage intensity and outdatedness) and a library update analysis (e.g., update intensity and delay) using 806 open-source projects. The two analyses aim to quantify usage and update practices holistically from the perspective of both open-source projects and third-party libraries. Then, we conduct a library risk analysis (e.g., potential risk and developer response) in terms of bugs with 15 popularly-used third-party libraries. This analysis aims to quantify the potential risk of using outdated libraries and the developer response to the risk. Our findings from the three analyses provide practical insights to developers and researchers on problems and potential solutions in maintaining third-party libraries (e.g., smart alerting and automated updating of outdated libraries). To demonstrate the usefulness of our findings, we propose a bug-driven alerting system for assisting developers to make confident decisions in updating third-party library versions. We have released our dataset to foster valuable applications and improve the ecosystem.

Phillip Rust, Bowen Shi, Skyler Wang et al.

A major impediment to the advancement of sign language translation (SLT) is data scarcity. Much of the sign language data currently available on the web cannot be used for training supervised models due to the lack of aligned captions. Furthermore, scaling SLT using large-scale web-scraped datasets bears privacy risks due to the presence of biometric information, which the responsible development of SLT technologies should account for. In this work, we propose a two-stage framework for privacy-aware SLT at scale that addresses both of these issues. We introduce SSVP-SLT, which leverages self-supervised video pretraining on anonymized and unannotated videos, followed by supervised SLT finetuning on a curated parallel dataset. SSVP-SLT achieves state-of-the-art finetuned and zero-shot gloss-free SLT performance on the How2Sign dataset, outperforming the strongest respective baselines by over 3 BLEU-4. Based on controlled experiments, we further discuss the advantages and limitations of self-supervised pretraining and anonymization via facial obfuscation for SLT.

HyoJung Han, Mohamed Anwar, Juan Pino et al. · meta-ai

Speech recognition and translation systems perform poorly on noisy inputs, which are frequent in realistic environments. Augmenting these systems with visual signals has the potential to improve robustness to noise. However, audio-visual (AV) data is only available in limited amounts and for fewer languages than audio-only resources. To address this gap, we present XLAVS-R, a cross-lingual audio-visual speech representation model for noise-robust speech recognition and translation in over 100 languages. It is designed to maximize the benefits of limited multilingual AV pre-training data, by building on top of audio-only multilingual pre-training and simplifying existing pre-training schemes. Extensive evaluation on the MuAViC benchmark shows the strength of XLAVS-R on downstream audio-visual speech recognition and translation tasks, where it outperforms the previous state of the art by up to 18.5% WER and 4.7 BLEU given noisy AV inputs, and enables strong zero-shot audio-visual ability with audio-only fine-tuning.

Ruifeng Yuan, Wanxing Chang, Weiwei Cao et al.

The evaluation of generated reports remains a critical challenge in Computed Tomography (CT) report generation, due to the large volume of text, the diversity and complexity of findings, and the presence of fine-grained, disease-oriented attributes. Conventional evaluation metrics offer only coarse measures of lexical overlap or entity matching and fail to reflect the granular diagnostic accuracy required for clinical use. To address this gap, we propose CT-FineBench, a benchmark built from CT-RATE and Merlin to evaluate the fine-grained factual consistency of CT reports, constructed from CT-RATE and Merlin. Our benchmark is constructed through a meticulous, Question-Answering (QA) based process: first, we identify and structure key, finding-specific clinical attributes (like location, size, margin). Second, we systematically transform these attributes into a QA dataset, where questions probe for specific clinical details grounded in gold-standard reports. The evaluation protocol for CT-FineBench involves using this QA dataset to query a machine-generated report and scoring the correctness of the answers. This allows for a comprehensive, interpretable, and clinically-relevant assessment, moving beyond superficial lexical overlap to pinpoint specific clinical errors. Experiments show that CT-FineBench correlates better with expert clinical assessment and is substantially more sensitive to fine-grained factual errors than prior metrics.

Xueyuan Chen, Shangzhe Li, Ruomei Liu et al.

Trapped by the label scarcity in molecular property prediction and drug design, graph contrastive learning (GCL) came forward. Leading contrastive learning works show two kinds of view generators, that is, random or learnable data corruption and domain knowledge incorporation. While effective, the two ways also lead to molecular semantics altering and limited generalization capability, respectively. To this end, we relate the \textbf{L}in\textbf{E} graph with \textbf{MO}lecular graph co\textbf{N}trastive learning and propose a novel method termed \textit{LEMON}. Specifically, by contrasting the given graph with the corresponding line graph, the graph encoder can freely encode the molecular semantics without omission. Furthermore, we present a new patch with edge attribute fusion and two local contrastive losses enhance information transmission and tackle hard negative samples. Compared with state-of-the-art (SOTA) methods for view generation, superior performance on molecular property prediction suggests the effectiveness of our proposed framework.

Guanghao Zheng, Bowen Shi, Mingxing Xu et al.

Vision encoders are indispensable for allowing impressive performance of Multi-modal Large Language Models (MLLMs) in vision language tasks such as visual question answering and reasoning. However, existing vision encoders focus on global image representations but overlook fine-grained regional analysis. They are limited in fine grained perception due to the scarcity of fine grained annotated data and the lack of a fine grained pre-training paradigm. In this paper, we propose GranViT, a novel Vision Transformer that integrates fine-grained feature extraction with semantic alignment to Large Language Models (LLMs) via region level autoregressive training. We first construct Gran-29M, a dataset comprising 2million natural and OCR images paired with over 180 million high-quality region-level annotations, to enable large scale fine grained pretraining. Consequently, we develop a pretraining-adaptation framework along with a self distillation mechanism to train fine-grained GranViT on Gran-29M. We sufficiently exploit the fine-grained annotations from Gran-29M to resort to bounding-box-to-caption regression to enhance localized visual representation of the vision encoder in the pretraining and caption-to-bounding-box regression to improve vision feature utilization and localization for LLM in the adaptation. We further incorporate a self distillation mechanism that imposes explicit localization constraints on the vision encoder to strengthen its regional reasoning capability. Extensive experiments show that GranViT surpasses existing vision encoders and attains strong transferability to varying LLMs. Remarkably, it achieves state-of-the-art results on fine-grained recognition, multimodal VQA, and OCR understanding.

Chuhan Zhang, Ye Zhang, Bowen Shi et al.

In deployment and application, large language models (LLMs) typically undergo safety alignment to prevent illegal and unethical outputs. However, the continuous advancement of jailbreak attack techniques, designed to bypass safety mechanisms with adversarial prompts, has placed increasing pressure on the security defenses of LLMs. Strengthening resistance to jailbreak attacks requires an in-depth understanding of the security mechanisms and vulnerabilities of LLMs. However, the vast number of parameters and complex structure of LLMs make analyzing security weaknesses from an internal perspective a challenging task. This paper presents NeuroBreak, a top-down jailbreak analysis system designed to analyze neuron-level safety mechanisms and mitigate vulnerabilities. We carefully design system requirements through collaboration with three experts in the field of AI security. The system provides a comprehensive analysis of various jailbreak attack methods. By incorporating layer-wise representation probing analysis, NeuroBreak offers a novel perspective on the model's decision-making process throughout its generation steps. Furthermore, the system supports the analysis of critical neurons from both semantic and functional perspectives, facilitating a deeper exploration of security mechanisms. We conduct quantitative evaluations and case studies to verify the effectiveness of our system, offering mechanistic insights for developing next-generation defense strategies against evolving jailbreak attacks.

Chung-Ming Chien, Andros Tjandra, Apoorv Vyas et al.

As the scale of generative models continues to grow, efficient reuse and adaptation of pre-trained models have become crucial considerations. In this work, we propose Voicebox Adapter, a novel approach that integrates fine-grained conditions into a pre-trained Voicebox speech generation model using a cross-attention module. To ensure a smooth integration of newly added modules with pre-trained ones, we explore various efficient fine-tuning approaches. Our experiment shows that the LoRA with bias-tuning configuration yields the best performance, enhancing controllability without compromising speech quality. Across three fine-grained conditional generation tasks, we demonstrate the effectiveness and resource efficiency of Voicebox Adapter. Follow-up experiments further highlight the robustness of Voicebox Adapter across diverse data setups.

Vineel Pratap, Andros Tjandra, Bowen Shi et al.

Expanding the language coverage of speech technology has the potential to improve access to information for many more people. However, current speech technology is restricted to about one hundred languages which is a small fraction of the over 7,000 languages spoken around the world. The Massively Multilingual Speech (MMS) project increases the number of supported languages by 10-40x, depending on the task. The main ingredients are a new dataset based on readings of publicly available religious texts and effectively leveraging self-supervised learning. We built pre-trained wav2vec 2.0 models covering 1,406 languages, a single multilingual automatic speech recognition model for 1,107 languages, speech synthesis models for the same number of languages, as well as a language identification model for 4,017 languages. Experiments show that our multilingual speech recognition model more than halves the word error rate of Whisper on 54 languages of the FLEURS benchmark while being trained on a small fraction of the labeled data.

Junran Wu, Xueyuan Chen, Bowen Shi et al.

In contrastive learning, the choice of ``view'' controls the information that the representation captures and influences the performance of the model. However, leading graph contrastive learning methods generally produce views via random corruption or learning, which could lead to the loss of essential information and alteration of semantic information. An anchor view that maintains the essential information of input graphs for contrastive learning has been hardly investigated. In this paper, based on the theory of graph information bottleneck, we deduce the definition of this anchor view; put differently, \textit{the anchor view with essential information of input graph is supposed to have the minimal structural uncertainty}. Furthermore, guided by structural entropy, we implement the anchor view, termed \textbf{SEGA}, for graph contrastive learning. We extensively validate the proposed anchor view on various benchmarks regarding graph classification under unsupervised, semi-supervised, and transfer learning and achieve significant performance boosts compared to the state-of-the-art methods.

Bowen Shi, Wei-Ning Hsu, Abdelrahman Mohamed

Audio-based automatic speech recognition (ASR) degrades significantly in noisy environments and is particularly vulnerable to interfering speech, as the model cannot determine which speaker to transcribe. Audio-visual speech recognition (AVSR) systems improve robustness by complementing the audio stream with the visual information that is invariant to noise and helps the model focus on the desired speaker. However, previous AVSR work focused solely on the supervised learning setup; hence the progress was hindered by the amount of labeled data available. In this work, we present a self-supervised AVSR framework built upon Audio-Visual HuBERT (AV-HuBERT), a state-of-the-art audio-visual speech representation learning model. On the largest available AVSR benchmark dataset LRS3, our approach outperforms prior state-of-the-art by ~50% (28.0% vs. 14.1%) using less than 10% of labeled data (433hr vs. 30hr) in the presence of babble noise, while reducing the WER of an audio-based model by over 75% (25.8% vs. 5.8%) on average.

Bowen Shi, Xiaopeng Zhang, Haohang Xu et al.

Collecting annotated data for semantic segmentation is time-consuming and hard to scale up. In this paper, we for the first time propose a unified framework, termed as Multi-Dataset Pretraining, to take full advantage of the fragmented annotations of different datasets. The highlight is that the annotations from different domains can be efficiently reused and consistently boost performance for each specific domain. This is achieved by first pretraining the network via the proposed pixel-to-prototype contrastive loss over multiple datasets regardless of their taxonomy labels, and followed by fine-tuning the pretrained model over specific dataset as usual. In order to better model the relationship among images and classes from different datasets, we extend the pixel level embeddings via cross dataset mixing and propose a pixel-to-class sparse coding strategy that explicitly models the pixel-class similarity over the manifold embedding space. In this way, we are able to increase intra-class compactness and inter-class separability, as well as considering inter-class similarity across different datasets for better transferability. Experiments conducted on several benchmarks demonstrate its superior performance. Notably, MDP consistently outperforms the pretrained models over ImageNet by a considerable margin, while only using less than 10% samples for pretraining.

Bowen Shi, Diane Brentari, Greg Shakhnarovich et al.

Fingerspelling, in which words are signed letter by letter, is an important component of American Sign Language. Most previous work on automatic fingerspelling recognition has assumed that the boundaries of fingerspelling regions in signing videos are known beforehand. In this paper, we consider the task of fingerspelling detection in raw, untrimmed sign language videos. This is an important step towards building real-world fingerspelling recognition systems. We propose a benchmark and a suite of evaluation metrics, some of which reflect the effect of detection on the downstream fingerspelling recognition task. In addition, we propose a new model that learns to detect fingerspelling via multi-task training, incorporating pose estimation and fingerspelling recognition (transcription) along with detection, and compare this model to several alternatives. The model outperforms all alternative approaches across all metrics, establishing a state of the art on the benchmark.

Chieh-Chi Kao, Bowen Shi, Ming Sun et al.

This paper proposes a network architecture mainly designed for audio tagging, which can also be used for weakly supervised acoustic event detection (AED). The proposed network consists of a modified DenseNet as the feature extractor, and a global average pooling (GAP) layer to predict frame-level labels at inference time. This architecture is inspired by the work proposed by Zhou et al., a well-known framework using GAP to localize visual objects given image-level labels. While most of the previous works on weakly supervised AED used recurrent layers with attention-based mechanism to localize acoustic events, the proposed network directly localizes events using the feature map extracted by DenseNet without any recurrent layers. In the audio tagging task of DCASE 2017, our method significantly outperforms the state-of-the-art method in F1 score by 5.3% on the dev set, and 6.0% on the eval set in terms of absolute values. For weakly supervised AED task in DCASE 2018, our model outperforms the state-of-the-art method in event-based F1 by 8.1% on the dev set, and 0.5% on the eval set in terms of absolute values, by using data augmentation and tri-training to leverage unlabeled data.

Bowen Shi, Shane Settle, Karen Livescu

Segmental models are sequence prediction models in which scores of hypotheses are based on entire variable-length segments of frames. We consider segmental models for whole-word ("acoustic-to-word") speech recognition, with the feature vectors defined using vector embeddings of segments. Such models are computationally challenging as the number of paths is proportional to the vocabulary size, which can be orders of magnitude larger than when using subword units like phones. We describe an efficient approach for end-to-end whole-word segmental models, with forward-backward and Viterbi decoding performed on a GPU and a simple segment scoring function that reduces space complexity. In addition, we investigate the use of pre-training via jointly trained acoustic word embeddings (AWEs) and acoustically grounded word embeddings (AGWEs) of written word labels. We find that word error rate can be reduced by a large margin by pre-training the acoustic segment representation with AWEs, and additional (smaller) gains can be obtained by pre-training the word prediction layer with AGWEs. Our final models improve over prior A2W models.

Shubham Toshniwal, Haoyue Shi, Bowen Shi et al.

Many natural language processing (NLP) tasks involve reasoning with textual spans, including question answering, entity recognition, and coreference resolution. While extensive research has focused on functional architectures for representing words and sentences, there is less work on representing arbitrary spans of text within sentences. In this paper, we conduct a comprehensive empirical evaluation of six span representation methods using eight pretrained language representation models across six tasks, including two tasks that we introduce. We find that, although some simple span representations are fairly reliable across tasks, in general the optimal span representation varies by task, and can also vary within different facets of individual tasks. We also find that the choice of span representation has a bigger impact with a fixed pretrained encoder than with a fine-tuned encoder.

Kaifeng Huang, Bihuan Chen, Bowen Shi et al.

As a mixed result of intensive dependency on third-party libraries, flexible mechanism to declare dependencies, and increased number of modules in a project, multiple versions of the same third-party library are directly depended in different modules of a project. Such library version inconsistencies can increase dependency maintenance cost, or even lead to dependency conflicts when modules are inter-dependent. Although automated build tools (e.g., Maven's enforcer plugin) provide partial support to detect library version inconsistencies, they do not provide any support to harmonize inconsistent library versions. We first conduct a survey with 131 Java developers from GitHub to retrieve first-hand information about the root causes, detection methods, reasons for fixing or not fixing, fixing strategies, fixing efforts, and tool expectations on library version inconsistencies. Then, based on the insights from our survey, we propose LibHarmo, an interactive, effort-aware library version harmonization technique, to detect library version inconsistencies, interactively suggest a harmonized version with the least harmonization efforts based on library API usage analysis, and refactor build configuration files. LibHarmo is currently developed for Java Maven projects. Our experimental study on 443 highly-starred Java Maven projects from GitHub indicates that i) LibHarmo identifies 621 library version inconsistencies covering 152 (34.3%) of projects, and ii) the average harmonization efforts are that 1 and 12 library API calls are affected, respectively due to the deleted and changed library APIs in the harmonized version. 5 library version inconsistencies have been confirmed, and 1 of them has been already harmonized by developers.