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.

We present VoiceCraft-Dub, a novel approach for automated video dubbing that synthesizes high-quality speech from text and facial cues. This task has broad applications in filmmaking, multimedia creation, and assisting voice-impaired individuals. Building on the success of Neural Codec Language Models (NCLMs) for speech synthesis, our method extends their capabilities by incorporating video features, ensuring that synthesized speech is time-synchronized and expressively aligned with facial movements while preserving natural prosody. To inject visual cues, we design adapters to align facial features with the NCLM token space and introduce audio-visual fusion layers to merge audio-visual information within the NCLM framework. Additionally, we curate CelebV-Dub, a new dataset of expressive, real-world videos specifically designed for automated video dubbing. Extensive experiments show that our model achieves high-quality, intelligible, and natural speech synthesis with accurate lip synchronization, outperforming existing methods in human perception and performing favorably in objective evaluations. We also adapt VoiceCraft-Dub for the video-to-speech task, demonstrating its versatility for various applications.

Language models require tokenized inputs. However, tokenization strategies for continuous data like audio and vision are often based on simple heuristics such as fixed sized convolutions or discrete clustering, which do not necessarily align with the semantic structure of the data. For speech in particular, the high resolution of waveforms (16,000 samples/second or more) presents a significant challenge as speech-based language models have had to use several times more tokens per word than text-based language models. In this work, we introduce a controllable self-supervised technique to merge speech representations into coarser syllable-like units while still preserving semantic information. We do this by 1) extracting noisy boundaries through analyzing correlations in pretrained encoder losses and 2) iteratively improving model representations with a novel distillation technique. Our method produces controllable-rate semantic units at as low as 5Hz and 60bps and achieves SotA in syllabic segmentation and clustering. Using these coarse tokens, we successfully train SyllableLM, a Speech Language Model (SpeechLM) that matches or outperforms current SotA SpeechLMs on a range of spoken language modeling tasks. SyllableLM also achieves significant improvements in efficiency with a 30x reduction in training compute and a 4x wall-clock inference speedup. Our code and checkpoints are available at https://www.github.com/alanbaade/SyllableLM

Multimodal foundation models, such as Gemini and ChatGPT, have revolutionized human-machine interactions by seamlessly integrating various forms of data. Developing a universal spoken language model that comprehends a wide range of natural language instructions is critical for bridging communication gaps and facilitating more intuitive interactions. However, the absence of a comprehensive evaluation benchmark poses a significant challenge. We present Dynamic-SUPERB Phase-2, an open and evolving benchmark for the comprehensive evaluation of instruction-based universal speech models. Building upon the first generation, this second version incorporates 125 new tasks contributed collaboratively by the global research community, expanding the benchmark to a total of 180 tasks, making it the largest benchmark for speech and audio evaluation. While the first generation of Dynamic-SUPERB was limited to classification tasks, Dynamic-SUPERB Phase-2 broadens its evaluation capabilities by introducing a wide array of novel and diverse tasks, including regression and sequence generation, across speech, music, and environmental audio. Evaluation results show that no model performed well universally. SALMONN-13B excelled in English ASR and Qwen2-Audio-7B-Instruct showed high accuracy in emotion recognition, but current models still require further innovations to handle a broader range of tasks. We open-source all task data and the evaluation pipeline at https://github.com/dynamic-superb/dynamic-superb.

Chart understanding presents a unique challenge for large vision-language models (LVLMs), as it requires the integration of sophisticated textual and visual reasoning capabilities. However, current LVLMs exhibit a notable imbalance between these skills, falling short on visual reasoning that is difficult to perform in text. We conduct a case study using a synthetic dataset solvable only through visual reasoning and show that model performance degrades significantly with increasing visual complexity, while human performance remains robust. We then introduce ChartMuseum, a new Chart Question Answering (QA) benchmark containing 1,162 expert-annotated questions spanning multiple reasoning types, curated from real-world charts across 184 sources, specifically built to evaluate complex visual and textual reasoning. Unlike prior chart understanding benchmarks---where frontier models perform similarly and near saturation---our benchmark exposes a substantial gap between model and human performance, while effectively differentiating model capabilities: although humans achieve 93% accuracy, the best-performing model Gemini-2.5-Pro attains only 63.0%, and the leading open-source LVLM Qwen2.5-VL-72B-Instruct achieves only 38.5%. Moreover, on questions requiring primarily visual reasoning, all models experience a 35%-55% performance drop from text-reasoning-heavy question performance. Lastly, our qualitative error analysis reveals specific categories of visual reasoning that are challenging for current LVLMs. Both ChartMuseum and the evaluation code are available at https://github.com/Liyan06/ChartMuseum.

RemovalFigure 1: TalkLess's interface lets creators edit their speech by skimming and browsing the outline pane to determine regions to cut, using the compression pane editing the transcript directly or by selecting a global-or paragraph-level compression amount, and listening to the original or generated audio aligned to the transcript in the audio pane.

We introduce RealSync, a novel dataset designed to significantly enhance the training and evaluation of models for audio-visual synchronization (AV Sync) tasks. Sourced from high-quality YouTube channels, RealSync covers a wide range of content domains, providing an improved scale, diversity, and alignment with broadcast content compared to existing datasets. It features extended-length video samples, catering to the critical need for more comprehensive, real-world training and evaluation materials. Alongside this dataset, we present StreamSync, a model tailored for real-world AV Sync applications. StreamSync is designed to be backbone agnostic and incorporates a streaming mechanism that processes consecutive video segments dynamically, iteratively refining synchronization predictions. This innovative approach enables StreamSync to outperform existing models, offering superior synchronization accuracy with minimal computational cost per iteration. Together, our dataset and the StreamSync model establish a new benchmark for AVSync research, promising to drive the development of more robust and practical AVSync methods. https://github.com/jvoas655/StreamSync

We introduce VOICECRAFT, a token infilling neural codec language model, that achieves state-of-the-art performance on both speech editing and zero-shot text-to-speech (TTS) on audiobooks, internet videos, and podcasts 1 .VOICECRAFT employs a Transformer decoder architecture and introduces a token rearrangement procedure that combines causal masking and delayed stacking to enable generation within an existing sequence.On speech editing tasks, VOICECRAFT produces edited speech that is nearly indistinguishable from unedited recordings in terms of naturalness, as evaluated by humans; for zero-shot TTS, our model outperforms prior SotA models including VALL-E and the popular commercial model XTTS v2.Crucially, the models are evaluated on challenging and realistic datasets, that consist of diverse accents, speaking styles, recording conditions, and background noise and music, and our model performs consistently well compared to other models and real recordings.In particular, for speech editing evaluation, we introduce a high quality, challenging, and realistic dataset named REALEDIT.We encourage readers to listen to the demos at https: //jasonppy.github.io/VoiceCraft_web.

Audio-visual representation learning aims to develop systems with human-like perception by utilizing correlation between auditory and visual information. However, current models often focus on a limited set of tasks, and generalization abilities of learned representations are unclear. To this end, we propose the AV-SUPERB benchmark that enables general-purpose evaluation of unimodal audio/visual and bimodal fusion representations on 7 datasets covering 5 audio-visual tasks in speech and audio processing. We evaluate 5 recent self-supervised models and show that none of these models generalize to all tasks, emphasizing the need for future study on improving universal model performance. In addition, we show that representations may be improved with intermediate-task fine-tuning and audio event classification with AudioSet serves as a strong intermediate task. We release our benchmark with evaluation code1and a model submission platform2to encourage further research in audio-visual learning.

Spatial sound reasoning is a fundamental human skill, enabling us to navigate and interpret our surroundings based on sound. In this paper we present BAT, which combines the spatial sound perception ability of a binaural acoustic scene analysis model with the natural language reasoning capabilities of a large language model (LLM) to replicate this innate ability. To address the lack of existing datasets of in-the-wild spatial sounds, we synthesized a binaural audio dataset using AudioSet and SoundSpaces 2.0. Next, we developed SpatialSoundQA, a spatial sound-based question-answering dataset, offering a range of QA tasks that train BAT in various aspects of spatial sound perception and reasoning. The acoustic front end encoder of BAT is a novel spatial audio encoder named Spatial Audio Spectrogram Transformer, or Spatial-AST, which by itself achieves strong performance across sound event detection, spatial localization, and distance estimation. By integrating Spatial-AST with LLaMA-2 7B model, BAT transcends standard Sound Event Localization and Detection (SELD) tasks, enabling the model to reason about the relationships between the sounds in its environment. Our experiments demonstrate BAT’s superior performance on both spatial sound perception and reasoning, showcasing the immense potential of LLMs in navigating and interpreting complex spatial audio environments.

We study phrase structure induction from visually-grounded speech. The core idea is to first segment the speech waveform into sequences of word segments, and subsequently induce phrase structure using the inferred segment-level continuous representations. We present the Audio-Visual Neural Syntax Learner (AV-NSL) that learns phrase structure by listening to audio and looking at images, without ever being exposed to text. By training on paired images and spoken captions, AV-NSL exhibits the capability to infer meaningful phrase structures that are comparable to those derived by naturally-supervised text parsers, for both English and German. Our findings extend prior work in unsupervised language acquisition from speech and grounded grammar induction, and present one approach to bridge the gap between the two topics.

In this paper, we show that representations capturing syllabic units emerge when training a self-supervised speech model with a visually-grounded training objective. We demonstrate that a nearly identical model architecture (HuBERT) trained with a masked language modeling loss does not exhibit this same ability, suggesting that the visual grounding objective is responsible for the emergence of this phenomenon. We propose the use of a minimum cut algorithm to automatically predict syllable boundaries in speech, followed by a 2-stage clustering method to group identical syllables together. We show that our model not only outperforms a state-of-the-art syllabic segmentation method on the language it was trained on (English), but also generalizes in a zero-shot fashion to Estonian. Finally, we show that the same model is capable of zero-shot generalization for a word segmentation task on 4 other languages from the Zerospeech Challenge, in some cases beating the previous state-of-the-art.

We present Fast-Slow Transformer for Visually Grounding Speech, or FaST-VGS. FaST-VGS is a Transformer-based model for learning the associations between raw speech waveforms and visual images. The model unifies dual-encoder and cross-attention architectures into a single model, reaping the superior retrieval speed of the former along with the accuracy of the latter. FaST-VGS achieves state-of-the-art speech-image retrieval accuracy on benchmark datasets, and its learned representations exhibit strong performance on the ZeroSpeech 2021 phonetic and semantic tasks.

In this paper, we propose a simple yet powerful improvement over the recent Self-Supervised Audio Spectrogram Transformer (SSAST) model for speech and audio classification.Specifically, we leverage the insight that the SSAST uses a very high masking ratio (75%) during pretraining, meaning that the vast majority of self-attention compute is performed on mask tokens.We address this by integrating the encoder-decoder architecture from Masked Autoencoders are Scalable Vision Learners (MAE) into the SSAST, where a deep encoder operates on only unmasked input, and a shallow decoder operates on encoder outputs and mask tokens.We find that MAE-like pretraining can provide a 3× speedup and 2× memory usage reduction over the vanilla SSAST using current audio pretraining strategies with ordinary model and input sizes.When finetuning on downstream tasks, which only uses the encoder, we find that our approach outperforms the SSAST on a variety of downstream tasks.We further conduct comprehensive evaluations into different strategies of pretraining and explore differences in MAE-style pretraining between the visual and audio domains.

OursFigure 1: HuBERT: sum of attention weights each frame receives from other frames.Ours (VG-HuBERT3): attention weights each frame receives from the [CLS A] token.Attention weights from different attention heads are coded with different colors.