David Harwath

Yuan Gong, Andrew Rouditchenko, Alexander H. Liu et al. · ibm-research, mit

In this paper, we first extend the recent Masked Auto-Encoder (MAE) model from a single modality to audio-visual multi-modalities. Subsequently, we propose the Contrastive Audio-Visual Masked Auto-Encoder (CAV-MAE) by combining contrastive learning and masked data modeling, two major self-supervised learning frameworks, to learn a joint and coordinated audio-visual representation. Our experiments show that the contrastive audio-visual correspondence learning objective not only enables the model to perform audio-visual retrieval tasks, but also helps the model learn a better joint representation. As a result, our fully self-supervised pretrained CAV-MAE achieves a new SOTA accuracy of 65.9% on VGGSound, and is comparable with the previous best supervised pretrained model on AudioSet in the audio-visual event classification task. Code and pretrained models are at https://github.com/yuangongnd/cav-mae.

Andrew Rouditchenko, Yung-Sung Chuang, Nina Shvetsova et al. · ibm-research, mit

Multilingual text-video retrieval methods have improved significantly in recent years, but the performance for other languages lags behind English. We propose a Cross-Lingual Cross-Modal Knowledge Distillation method to improve multilingual text-video retrieval. Inspired by the fact that English text-video retrieval outperforms other languages, we train a student model using input text in different languages to match the cross-modal predictions from teacher models using input text in English. We propose a cross entropy based objective which forces the distribution over the student's text-video similarity scores to be similar to those of the teacher models. We introduce a new multilingual video dataset, Multi-YouCook2, by translating the English captions in the YouCook2 video dataset to 8 other languages. Our method improves multilingual text-video retrieval performance on Multi-YouCook2 and several other datasets such as Multi-MSRVTT and VATEX. We also conducted an analysis on the effectiveness of different multilingual text models as teachers. The code, models, and dataset are available at https://github.com/roudimit/c2kd.

Yuan Tseng, Layne Berry, Yi-Ting Chen et al. · meta-ai, mit

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 code and a model submission platform to encourage further research in audio-visual learning.

Yi-Jen Shih, Hsuan-Fu Wang, Heng-Jui Chang et al. · mit

Data-driven speech processing models usually perform well with a large amount of text supervision, but collecting transcribed speech data is costly. Therefore, we propose SpeechCLIP, a novel framework bridging speech and text through images to enhance speech models without transcriptions. We leverage state-of-the-art pre-trained HuBERT and CLIP, aligning them via paired images and spoken captions with minimal fine-tuning. SpeechCLIP outperforms prior state-of-the-art on image-speech retrieval and performs zero-shot speech-text retrieval without direct supervision from transcriptions. Moreover, SpeechCLIP can directly retrieve semantically related keywords from speech.

Anuj Diwan, Layne Berry, Eunsol Choi et al.

Recent visuolinguistic pre-trained models show promising progress on various end tasks such as image retrieval and video captioning. Yet, they fail miserably on the recently proposed Winoground dataset, which challenges models to match paired images and English captions, with items constructed to overlap lexically but differ in meaning (e.g., "there is a mug in some grass" vs. "there is some grass in a mug"). By annotating the dataset using new fine-grained tags, we show that solving the Winoground task requires not just compositional language understanding, but a host of other abilities like commonsense reasoning or locating small, out-of-focus objects in low-resolution images. In this paper, we identify the dataset's main challenges through a suite of experiments on related tasks (probing task, image retrieval task), data augmentation, and manual inspection of the dataset. Our analysis suggests that a main challenge in visuolinguistic models may lie in fusing visual and textual representations, rather than in compositional language understanding. We release our annotation and code at https://github.com/ajd12342/why-winoground-hard .

Layne Berry, Yi-Jen Shih, Hsuan-Fu Wang et al. · mit

This work investigates the use of large-scale, English-only pre-trained models (CLIP and HuBERT) for multilingual image-speech retrieval. For non-English image-speech retrieval, we outperform the current state-of-the-art performance by a wide margin both when training separate models for each language, and with a single model which processes speech in all three languages. We identify key differences in model behavior and performance between English and non-English settings, attributable to the English-only pre-training of CLIP and HuBERT, and investigate how fine-tuning the pre-trained models impacts these differences. Finally, we show that our models can be used for mono- and cross-lingual speech-text retrieval and cross-lingual speech-speech retrieval, despite never having seen any parallel speech-text or speech-speech data during training.

Puyuan Peng, David Harwath

We present a method for visually-grounded spoken term discovery. After training either a HuBERT or wav2vec2.0 model to associate spoken captions with natural images, we show that powerful word segmentation and clustering capability emerges within the model's self-attention heads. Our experiments reveal that this ability is not present to nearly the same extent in the base HuBERT and wav2vec2.0 models, suggesting that the visual grounding task is a crucial component of the word discovery capability we observe. We also evaluate our method on the Buckeye word segmentation and ZeroSpeech spoken term discovery tasks, where we perform on par with or better than currently published methods on several metrics. Code and model weights are available at https://github.com/jasonppy/word-discovery.

Cheng-I Jeff Lai, Freda Shi, Puyuan Peng et al. · mit

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.

Anuj Diwan, Eunsol Choi, David Harwath

We present the first unified study of the efficiency of self-attention-based Transformer variants spanning text, speech and vision. We identify input length thresholds (tipping points) at which efficient Transformer variants become more efficient than vanilla models, using a variety of efficiency metrics (latency, throughput, and memory). To conduct this analysis for speech, we introduce L-HuBERT, a novel local-attention variant of a self-supervised speech model. We observe that these thresholds are (a) much higher than typical dataset sequence lengths and (b) dependent on the metric and modality, showing that choosing the right model depends on modality, task type (long-form vs. typical context) and resource constraints (time vs. memory). By visualising the breakdown of the computational costs for transformer components, we also show that non-self-attention components exhibit significant computational costs. We release our profiling toolkit at https://github.com/ajd12342/profiling-transformers .

Anuj Diwan, Eunsol Choi, David Harwath

We introduce ParaSpeechCLAP, a dual-encoder contrastive model that maps speech and text style captions into a common embedding space, supporting a wide range of intrinsic (speaker-level) and situational (utterance-level) descriptors (such as pitch, texture and emotion) far beyond the narrow set handled by existing models. We train specialized ParaSpeechCLAP-Intrinsic and ParaSpeechCLAP-Situational models alongside a unified ParaSpeechCLAP-Combined model, finding that specialization yields stronger performance on individual style dimensions while the unified model excels on compositional evaluation. We further show that ParaSpeechCLAP-Intrinsic benefits from an additional classification loss and class-balanced training. We demonstrate our models' performance on style caption retrieval, speech attribute classification and as an inference-time reward model that improves style-prompted TTS without additional training. ParaSpeechCLAP outperforms baselines on most metrics across all three applications. Our models and code are released at https://github.com/ajd12342/paraspeechclap .

Yi-Jen Shih, Zoi Gkalitsiou, Alexandros G. Dimakis et al.

Clinical diagnosis of stuttering requires an assessment by a licensed speech-language pathologist. However, this process is time-consuming and requires clinicians with training and experience in stuttering and fluency disorders. Unfortunately, only a small percentage of speech-language pathologists report being comfortable working with individuals who stutter, which is inadequate to accommodate for the 80 million individuals who stutter worldwide. Developing machine learning models for detecting stuttered speech would enable universal and automated screening for stuttering, enabling speech pathologists to identify and follow up with patients who are most likely to be diagnosed with a stuttering speech disorder. Previous research in this area has predominantly focused on utterance-level detection, which is not sufficient for clinical settings where word-level annotation of stuttering is the norm. In this study, we curated a stuttered speech dataset with word-level annotations and introduced a word-level stuttering speech detection model leveraging self-supervised speech models. Our evaluation demonstrates that our model surpasses previous approaches in word-level stuttering speech detection. Additionally, we conducted an extensive ablation analysis of our method, providing insight into the most important aspects of adapting self-supervised speech models for stuttered speech detection.

Anuj Diwan, Ching-Feng Yeh, Wei-Ning Hsu et al.

Automatic speech recognition research focuses on training and evaluating on static datasets. Yet, as speech models are increasingly deployed on personal devices, such models encounter user-specific distributional shifts. To simulate this real-world scenario, we introduce LibriContinual, a continual learning benchmark for speaker-specific domain adaptation derived from LibriVox audiobooks, with data corresponding to 118 individual speakers and 6 train splits per speaker of different sizes. Additionally, current speech recognition models and continual learning algorithms are not optimized to be compute-efficient. We adapt a general-purpose training algorithm NetAug for ASR and create a novel Conformer variant called the DisConformer (Disentangled Conformer). This algorithm produces ASR models consisting of a frozen 'core' network for general-purpose use and several tunable 'augment' networks for speaker-specific tuning. Using such models, we propose a novel compute-efficient continual learning algorithm called DisentangledCL. Our experiments show that the DisConformer models significantly outperform baselines on general ASR i.e. LibriSpeech (15.58% rel. WER on test-other). On speaker-specific LibriContinual they significantly outperform trainable-parameter-matched baselines (by 20.65% rel. WER on test) and even match fully finetuned baselines in some settings.

Chiori Hori, Puyuan Peng, David Harwath et al.

To realize human-robot collaboration, robots need to execute actions for new tasks according to human instructions given finite prior knowledge. Human experts can share their knowledge of how to perform a task with a robot through multi-modal instructions in their demonstrations, showing a sequence of short-horizon steps to achieve a long-horizon goal. This paper introduces a method for robot action sequence generation from instruction videos using (1) an audio-visual Transformer that converts audio-visual features and instruction speech to a sequence of robot actions called dynamic movement primitives (DMPs) and (2) style-transfer-based training that employs multi-task learning with video captioning and weakly-supervised learning with a semantic classifier to exploit unpaired video-action data. We built a system that accomplishes various cooking actions, where an arm robot executes a DMP sequence acquired from a cooking video using the audio-visual Transformer. Experiments with Epic-Kitchen-100, YouCookII, QuerYD, and in-house instruction video datasets show that the proposed method improves the quality of DMP sequences by 2.3 times the METEOR score obtained with a baseline video-to-action Transformer. The model achieved 32% of the task success rate with the task knowledge of the object.

Abner Hernandez, Eunjung Yeo, Kwanghee Choi et al.

The limited availability of dysarthric speech data makes cross-lingual detection an important but challenging problem. A key difficulty is that speech representations often encode language-dependent structure that can confound dysarthria detection. We propose a representation-level language shift (LS) that aligns source-language self-supervised speech representations with the target-language distribution using centroid-based vector adaptation estimated from healthy-control speech. We evaluate the approach on oral DDK recordings from Parkinson's disease speech datasets in Czech, German, and Spanish under both cross-lingual and multilingual settings. LS substantially improves sensitivity and F1 in cross-lingual settings, while yielding smaller but consistent gains in multilingual settings. Representation analysis further shows that LS reduces language identity in the embedding space, supporting the interpretation that LS removes language-dependent structure.

Luke Strgar, David Harwath

We apply transfer learning to the task of phoneme segmentation and demonstrate the utility of representations learned in self-supervised pre-training for the task. Our model extends transformer-style encoders with strategically placed convolutions that manipulate features learned in pre-training. Using the TIMIT and Buckeye corpora we train and test the model in the supervised and unsupervised settings. The latter case is accomplished by furnishing a noisy label-set with the predictions of a separate model, it having been trained in an unsupervised fashion. Results indicate our model eclipses previous state-of-the-art performance in both settings and on both datasets. Finally, following observations during published code review and attempts to reproduce past segmentation results, we find a need to disambiguate the definition and implementation of widely-used evaluation metrics. We resolve this ambiguity by delineating two distinct evaluation schemes and describing their nuances.

Reem Gody, David Harwath

Self-supervised learning (SSL) has been able to leverage unlabeled data to boost the performance of automatic speech recognition (ASR) models when we have access to only a small amount of transcribed speech data. However, this raises the question of which subset of the available unlabeled data should be selected for transcription. Our work investigates different unsupervised data selection techniques for fine-tuning the HuBERT model under a limited transcription budget. We investigate the impact of speaker diversity, gender bias, and topic diversity on the downstream ASR performance. We also devise two novel techniques for unsupervised data selection: pre-training loss based data selection and the perplexity of byte pair encoded clustered units (PBPE) and we show how these techniques compare to pure random data selection. Finally, we analyze the correlations between the inherent characteristics of the selected fine-tuning subsets as well as how these characteristics correlate with the resultant word error rate. We demonstrate the importance of token diversity, speaker diversity, and topic diversity in achieving the best performance in terms of WER.

Chien-yu Huang, Wei-Chih Chen, Shu-wen Yang et al. · cmu, mit

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.

Wei-Cheng Tseng, Yi-Jen Shih, David Harwath et al.

Audio-visual pre-trained models have gained substantial attention recently and demonstrated superior performance on various audio-visual tasks. This study investigates whether pre-trained audio-visual models demonstrate non-arbitrary associations between sounds and visual representations$\unicode{x2013}$known as sound symbolism$\unicode{x2013}$which is also observed in humans. We developed a specialized dataset with synthesized images and audio samples and assessed these models using a non-parametric approach in a zero-shot setting. Our findings reveal a significant correlation between the models' outputs and established patterns of sound symbolism, particularly in models trained on speech data. These results suggest that such models can capture sound-meaning connections akin to human language processing, providing insights into both cognitive architectures and machine learning strategies.

Kim Sung-Bin, Joohyun Chang, David Harwath et al.

Talking face editing and face generation have often been studied as distinct problems. In this work, we propose viewing both not as separate tasks but as subtasks of a unifying formulation, speech-conditional facial motion infilling. We explore facial motion infilling as a self-supervised pretext task that also serves as a unifying formulation of dynamic talking face synthesis. To instantiate this idea, we propose FacEDiT, a speech-conditional Diffusion Transformer trained with flow matching. Inspired by masked autoencoders, FacEDiT learns to synthesize masked facial motions conditioned on surrounding motions and speech. This formulation enables both localized generation and edits, such as substitution, insertion, and deletion, while ensuring seamless transitions with unedited regions. In addition, biased attention and temporal smoothness constraints enhance boundary continuity and lip synchronization. To address the lack of a standard editing benchmark, we introduce FacEDiTBench, the first dataset for talking face editing, featuring diverse edit types and lengths, along with new evaluation metrics. Extensive experiments validate that talking face editing and generation emerge as subtasks of speech-conditional motion infilling; FacEDiT produces accurate, speech-aligned facial edits with strong identity preservation and smooth visual continuity while generalizing effectively to talking face generation.

Zhisheng Zheng, Puyuan Peng, Anuj Diwan et al.

We introduce VoiceCraft-X, an autoregressive neural codec language model which unifies multilingual speech editing and zero-shot Text-to-Speech (TTS) synthesis across 11 languages: English, Mandarin, Korean, Japanese, Spanish, French, German, Dutch, Italian, Portuguese, and Polish. VoiceCraft-X utilizes the Qwen3 large language model for phoneme-free cross-lingual text processing and a novel token reordering mechanism with time-aligned text and speech tokens to handle both tasks as a single sequence generation problem. The model generates high-quality, natural-sounding speech, seamlessly creating new audio or editing existing recordings within one framework. VoiceCraft-X shows robust performance in diverse linguistic settings, even with limited per-language data, underscoring the power of unified autoregressive approaches for advancing complex, real-world multilingual speech applications. Audio samples are available at https://zhishengzheng.com/voicecraft-x/.

Yi-Jen Shih, David Harwath

Speech Foundation Models have gained significant attention recently. Prior works have shown that the fusion of representations from multiple layers of the same model or the fusion of multiple models can improve performance on downstream tasks. We unify these two fusion strategies by proposing an interface module that enables fusion across multiple upstream speech models while integrating information across their layers. We conduct extensive experiments on different self-supervised and supervised models across various speech tasks, including ASR and paralinguistic analysis, and demonstrate that our method outperforms prior fusion approaches. We further analyze its scalability concerning model size and count, highlighting the importance of selecting appropriate upstream models. Our results show that the proposed interface provides an additional performance boost when given a suitable upstream model selection, making it a promising approach for utilizing Speech Foundation Models.

Anuj Diwan, Zhisheng Zheng, David Harwath et al.

We introduce Paralinguistic Speech Captions (ParaSpeechCaps), a large-scale dataset that annotates speech utterances with rich style captions. While rich abstract tags (e.g. guttural, nasal, pained) have been explored in small-scale human-annotated datasets, existing large-scale datasets only cover basic tags (e.g. low-pitched, slow, loud). We combine off-the-shelf text and speech embedders, classifiers and an audio language model to automatically scale rich tag annotations for the first time. ParaSpeechCaps covers a total of 59 style tags, including both speaker-level intrinsic tags and utterance-level situational tags. It consists of 342 hours of human-labelled data (PSC-Base) and 2427 hours of automatically annotated data (PSC-Scaled). We finetune Parler-TTS, an open-source style-prompted TTS model, on ParaSpeechCaps, and achieve improved style consistency (+7.9% Consistency MOS) and speech quality (+15.5% Naturalness MOS) over the best performing baseline that combines existing rich style tag datasets. We ablate several of our dataset design choices to lay the foundation for future work in this space. Our dataset, models and code are released at https://github.com/ajd12342/paraspeechcaps .

Kwanghee Choi, Eunjung Yeo, Cheol Jun Cho et al.

Self-supervised speech models (S3Ms) are known to encode rich phonetic information, yet how this information is structured remains underexplored. We conduct a comprehensive study across 96 languages to analyze the underlying structure of S3M representations, with particular attention to phonological vectors. We first show that there exist linear directions within the model's representation space that correspond to phonological features. We further demonstrate that the scale of these phonological vectors correlate to the degree of acoustic realization of their corresponding phonological features in a continuous manner. For example, the difference between [d] and [t] yields a voicing vector: adding this vector to [p] produces [b], while scaling it results in a continuum of voicing. Together, these findings indicate that S3Ms encode speech using phonologically interpretable and compositional vectors, demonstrating phonological vector arithmetic. All code and interactive demos are available at https://github.com/juice500ml/phonetic-arithmetic .

Puyuan Peng, Po-Yao Huang, Shang-Wen Li et al. · meta-ai, mit

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. 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 VALLE 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.

Puyuan Peng, Brian Yan, Shinji Watanabe et al.

We investigate the emergent abilities of the recently proposed web-scale speech model Whisper, by adapting it to unseen tasks with prompt engineering. We selected three tasks: audio-visual speech recognition (AVSR), code-switched speech recognition (CS-ASR), and speech translation (ST) on unseen language pairs. We design task-specific prompts, by either leveraging another large-scale model, or simply manipulating the special tokens in the default prompts. Experiments show that compared to the default prompts, our proposed prompts improve performance by 10% to 45% on the three zero-shot tasks, and even outperform SotA supervised models on some datasets. In addition, our experiments reveal many interesting properties of Whisper, including its robustness to prompts, bias on accents, and the multilingual understanding in its latent space. Code is available at https://github.com/jasonppy/PromptingWhisper

Zhisheng Zheng, Puyuan Peng, Ziyang Ma et al.

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.

Changan Chen, Kumar Ashutosh, Rohit Girdhar et al. · meta-ai

We propose a novel self-supervised embedding to learn how actions sound from narrated in-the-wild egocentric videos. Whereas existing methods rely on curated data with known audio-visual correspondence, our multimodal contrastive-consensus coding (MC3) embedding reinforces the associations between audio, language, and vision when all modality pairs agree, while diminishing those associations when any one pair does not. We show our approach can successfully discover how the long tail of human actions sound from egocentric video, outperforming an array of recent multimodal embedding techniques on two datasets (Ego4D and EPIC-Sounds) and multiple cross-modal tasks.

Hsuan-Fu Wang, Yi-Jen Shih, Heng-Jui Chang et al. · mit

The recently proposed visually grounded speech model SpeechCLIP is an innovative framework that bridges speech and text through images via CLIP without relying on text transcription. On this basis, this paper introduces two extensions to SpeechCLIP. First, we apply the Continuous Integrate-and-Fire (CIF) module to replace a fixed number of CLS tokens in the cascaded architecture. Second, we propose a new hybrid architecture that merges the cascaded and parallel architectures of SpeechCLIP into a multi-task learning framework. Our experimental evaluation is performed on the Flickr8k and SpokenCOCO datasets. The results show that in the speech keyword extraction task, the CIF-based cascaded SpeechCLIP model outperforms the previous cascaded SpeechCLIP model using a fixed number of CLS tokens. Furthermore, through our hybrid architecture, cascaded task learning boosts the performance of the parallel branch in image-speech retrieval tasks.

Shih-Heng Wang, Zih-Ching Chen, Jiatong Shi et al. · meta-ai, mit

The utilization of speech Self-Supervised Learning (SSL) models achieves impressive performance on Automatic Speech Recognition (ASR). However, in low-resource language ASR, they encounter the domain mismatch problem between pre-trained and low-resource languages. Typical solutions like fine-tuning the SSL model suffer from high computation costs while using frozen SSL models as feature extractors comes with poor performance. To handle these issues, we extend a conventional efficient fine-tuning scheme based on the adapter. We add an extra intermediate adaptation to warm up the adapter and downstream model initialization. Remarkably, we update only 1-5% of the total model parameters to achieve the adaptation. Experimental results on the ML-SUPERB dataset show that our solution outperforms conventional efficient fine-tuning. It achieves up to a 28% relative improvement in the Character/Phoneme error rate when adapting to unseen languages.

Kwanghee Choi, Eunjung Yeo, Cheol Jun Cho et al.

Transformer-based self-supervised speech models (S3Ms) are often described as contextualized, yet what this entails remains unclear. Here, we focus on how a single frame-level S3M representation can encode phones and their surrounding context. Prior work has shown that S3Ms represent phones compositionally; for example, phonological vectors such as voicing, bilabiality, and nasality vectors are superposed in the S3M representation of [m]. We extend this view by proposing that phonological information from a sequence of neighboring phones is also compositionally encoded in a single frame, such that vectors corresponding to previous, current, and next phones are superposed within a single frame-level representation. We show that this structure has several properties, including orthogonality between relative positions, and emergence of implicit phonetic boundaries. Together, our findings advance our understanding of context-dependent S3M representations.

Kim Sung-Bin, Jeongsoo Choi, Puyuan Peng et al.

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.

Hung-Chieh Fang, Nai-Xuan Ye, Yi-Jen Shih et al.

Recent advances in self-supervised speech models have shown significant improvement in many downstream tasks. However, these models predominantly centered on frame-level training objectives, which can fall short in spoken language understanding tasks that require semantic comprehension. Existing works often rely on additional speech-text data as intermediate targets, which is costly in the real-world setting. To address this challenge, we propose Pseudo-Word HuBERT (PW-HuBERT), a framework that integrates pseudo word-level targets into the training process, where the targets are derived from a visually-ground speech model, notably eliminating the need for speech-text paired data. Our experimental results on four spoken language understanding (SLU) benchmarks suggest the superiority of our model in capturing semantic information.

Wei-Cheng Tseng, David Harwath

Recent advancements in neural audio codecs have not only enabled superior audio compression but also enhanced speech synthesis techniques. Researchers are now exploring their potential as universal acoustic feature extractors for a broader range of speech processing tasks. Building on this trend, we introduce Codec2Vec, the first speech representation learning framework that relies exclusively on discrete audio codec units. This approach offers several advantages, including improved data storage and transmission efficiency, faster training, and enhanced data privacy. We explore masked prediction with various training target derivation strategies to thoroughly understand the effectiveness of this framework. Evaluated on the SUPERB benchmark, Codec2Vec achieves competitive performance compared to continuous-input models while reducing storage requirements by up to 16.5x and training time by 2.3x, showcasing its scalability and efficiency.

Zhisheng Zheng, Xiaohang Sun, Tuan Dinh et al.

End-to-end speech-to-speech translation (S2ST) systems typically struggle with a critical data bottleneck: the scarcity of parallel speech-to-speech corpora. To overcome this, we introduce RosettaSpeech, a novel zero-shot framework trained exclusively on monolingual speech-text data augmented by machine translation supervision. Unlike prior works that rely on complex cascaded pseudo-labeling, our approach strategically utilizes text as a semantic bridge during training to synthesize translation targets, thereby eliminating the need for parallel speech pairs while maintaining a direct, end-to-end inference pipeline. Empirical evaluations on the CVSS-C benchmark demonstrate that RosettaSpeech achieves state-of-the-art zero-shot performance, surpassing leading baselines by significant margins - achieving ASR-BLEU scores of 25.17 for German-to-English (+27% relative gain) and 29.86 for Spanish-to-English (+14%). Crucially, our model effectively preserves the source speaker's voice without ever seeing paired speech data. We further analyze the impact of data scaling and demonstrate the model's capability in many-to-one translation, offering a scalable solution for extending high-quality S2ST to "text-rich, speech-poor" languages.

Liming Wang, Junrui Ni, Kai-Wei Chang et al.

Training speech recognizers with unpaired speech and text -- known as unsupervised speech recognition (UASR) -- is a crucial step toward extending ASR to low-resource languages in the long-tail distribution and enabling multimodal learning from non-parallel data. However, existing approaches based on phones often rely on costly resources such as grapheme-to-phoneme converters (G2Ps) and struggle to generalize to languages with ambiguous phoneme boundaries due to training instability. In this paper, we address both challenges by introducing a syllable-level UASR framework based on masked language modeling, which avoids the need for G2P and the instability of GAN-based methods. Our approach achieves up to a 40\% relative reduction in character error rate (CER) on LibriSpeech and generalizes effectively to Mandarin, a language that has remained particularly difficult for prior methods. Code will be released upon acceptance.

Younghan Park, Anuj Diwan, David Harwath et al.

Podcasts have become daily companions for half a billion users. Given the enormous amount of podcast content available, highlights provide a valuable signal that helps viewers get the gist of an episode and decide if they want to invest in listening to it in its entirety. However, identifying highlights automatically is challenging due to the unstructured and long-form nature of the content. We introduce Rhapsody, a dataset of 13K podcast episodes paired with segment-level highlight scores derived from YouTube's 'most replayed' feature. We frame the podcast highlight detection as a segment-level binary classification task. We explore various baseline approaches, including zero-shot prompting of language models and lightweight fine-tuned language models using segment-level classification heads. Our experimental results indicate that even state-of-the-art language models like GPT-4o and Gemini struggle with this task, while models fine-tuned with in-domain data significantly outperform their zero-shot performance. The fine-tuned model benefits from leveraging both speech signal features and transcripts. These findings highlight the challenges for fine-grained information access in long-form spoken media.

Yi-Jen Shih, David Harwath

Self-supervised speech (SSL) models have recently become widely adopted for many downstream speech processing tasks. The general usage pattern is to employ SSL models as feature extractors, and then train a downstream prediction head to solve a specific task. However, different layers of SSL models have been shown to capture different types of information, and the methods of combining them are not well studied. To this end, we extend the general framework for SSL model utilization by proposing the interface that connects the upstream and downstream. Under this view, the dominant technique of combining features via a layerwise weighted sum can be regarded as a specific interface. We propose several alternative interface designs and demonstrate that the weighted sum interface is suboptimal for many tasks. In particular, we show that a convolutional interface whose depth scales logarithmically with the depth of the upstream model consistently outperforms many other interface designs.

Changan Chen, Puyuan Peng, Ami Baid et al.

Generating realistic audio for human actions is important for many applications, such as creating sound effects for films or virtual reality games. Existing approaches implicitly assume total correspondence between the video and audio during training, yet many sounds happen off-screen and have weak to no correspondence with the visuals -- resulting in uncontrolled ambient sounds or hallucinations at test time. We propose a novel ambient-aware audio generation model, AV-LDM. We devise a novel audio-conditioning mechanism to learn to disentangle foreground action sounds from the ambient background sounds in in-the-wild training videos. Given a novel silent video, our model uses retrieval-augmented generation to create audio that matches the visual content both semantically and temporally. We train and evaluate our model on two in-the-wild egocentric video datasets, Ego4D and EPIC-KITCHENS, and we introduce Ego4D-Sounds -- 1.2M curated clips with action-audio correspondence. Our model outperforms an array of existing methods, allows controllable generation of the ambient sound, and even shows promise for generalizing to computer graphics game clips. Overall, our approach is the first to focus video-to-audio generation faithfully on the observed visual content despite training from uncurated clips with natural background sounds.

Jordan Voas, Raymond Mooney, David Harwath

We introduce Semantic Parsing in Contextual Environments (SPICE), a task designed to enhance artificial agents' contextual awareness by integrating multimodal inputs with prior contexts. SPICE goes beyond traditional semantic parsing by offering a structured, interpretable framework for dynamically updating an agent's knowledge with new information, mirroring the complexity of human communication. We develop the VG-SPICE dataset, crafted to challenge agents with visual scene graph construction from spoken conversational exchanges, highlighting speech and visual data integration. We also present the Audio-Vision Dialogue Scene Parser (AViD-SP) developed for use on VG-SPICE. These innovations aim to improve multimodal information processing and integration. Both the VG-SPICE dataset and the AViD-SP model are publicly available.

Anuj Diwan, Anirudh Srinivasan, David Harwath et al.

Existing speech-to-speech translation (S2ST) models fall into two camps: they either leverage text as an intermediate step or require hundreds of hours of parallel speech data. Both approaches are incompatible with textless languages or language pairs with limited parallel data. We present PFB, a framework for training textless S2ST models that require just dozens of hours of parallel speech data. We first pretrain a model on large-scale monolingual speech data, finetune it with a small amount of parallel speech data (20-60 hours), and lastly train with an unsupervised backtranslation objective. We train and evaluate our models for English-to-German, German-to-English and Marathi-to-English translation on three different domains (European Parliament, Common Voice, and All India Radio) with single-speaker synthesized speech. Evaluated using the ASR-BLEU metric, our models achieve reasonable performance on all three domains, with some being within 1-2 points of our higher-resourced topline.

Andrew Rouditchenko, Sameer Khurana, Samuel Thomas et al.

Recent models such as XLS-R and Whisper have made multilingual speech technologies more accessible by pre-training on audio from around 100 spoken languages each. However, there are thousands of spoken languages worldwide, and adapting to new languages is an important problem. In this work, we aim to understand which model adapts better to languages unseen during pre-training. We fine-tune both models on 13 unseen languages and 18 seen languages. Our results show that the number of hours seen per language and language family during pre-training is predictive of how the models compare, despite the significant differences in the pre-training methods.

Puyuan Peng, Shang-Wen Li, Okko Räsänen et al.

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.

Alan Baade, Puyuan Peng, David Harwath

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 3x speedup and 2x memory usage reduction over the vanilla SSAST using current audio pretraining strategies with ordinary model and input sizes. When fine-tuning 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.

Puyuan Peng, David Harwath

In this paper, we describe our submissions to the ZeroSpeech 2021 Challenge and SUPERB benchmark. Our submissions are based on the recently proposed FaST-VGS model, which is a Transformer-based model that learns to associate raw speech waveforms with semantically related images, all without the use of any transcriptions of the speech. Additionally, we introduce a novel extension of this model, FaST-VGS+, which is learned in a multi-task fashion with a masked language modeling objective in addition to the visual grounding objective. On ZeroSpeech 2021, we show that our models perform competitively on the ABX task, outperform all other concurrent submissions on the Syntactic and Semantic tasks, and nearly match the best system on the Lexical task. On the SUPERB benchmark, we show that our models also achieve strong performance, in some cases even outperforming the popular wav2vec2.0 model.

Nina Shvetsova, Brian Chen, Andrew Rouditchenko et al.

Multi-modal learning from video data has seen increased attention recently as it allows to train semantically meaningful embeddings without human annotation enabling tasks like zero-shot retrieval and classification. In this work, we present a multi-modal, modality agnostic fusion transformer approach that learns to exchange information between multiple modalities, such as video, audio, and text, and integrate them into a joined multi-modal representation to obtain an embedding that aggregates multi-modal temporal information. We propose to train the system with a combinatorial loss on everything at once, single modalities as well as pairs of modalities, explicitly leaving out any add-ons such as position or modality encoding. At test time, the resulting model can process and fuse any number of input modalities. Moreover, the implicit properties of the transformer allow to process inputs of different lengths. To evaluate the proposed approach, we train the model on the large scale HowTo100M dataset and evaluate the resulting embedding space on four challenging benchmark datasets obtaining state-of-the-art results in zero-shot video retrieval and zero-shot video action localization.

Kevin Duarte, Brian Chen, Nina Shvetsova et al.

The task of multimodal learning has seen a growing interest recently as it allows for training neural architectures based on different modalities such as vision, text, and audio. One challenge in training such models is that they need to jointly learn semantic concepts and their relationships across different input representations. Capsule networks have been shown to perform well in context of capturing the relation between low-level input features and higher-level concepts. However, capsules have so far mainly been used only in small-scale fully supervised settings due to the resource demand of conventional routing algorithms. We present a new multimodal capsule network that allows us to leverage the strength of capsules in the context of a multimodal learning framework on large amounts of video data. To adapt the capsules to large-scale input data, we propose a novel routing by self-attention mechanism that selects relevant capsules which are then used to generate a final joint multimodal feature representation. This allows not only for robust training with noisy video data, but also to scale up the size of the capsule network compared to traditional routing methods while still being computationally efficient. We evaluate the proposed architecture by pretraining it on a large-scale multimodal video dataset and applying it on four datasets in two challenging downstream tasks. Results show that the proposed multimodal capsule network is not only able to improve results compared to other routing techniques, but also achieves competitive performance on the task of multimodal learning.

Andrew Rouditchenko, Angie Boggust, David Harwath et al.

In this paper, we explore self-supervised audio-visual models that learn from instructional videos. Prior work has shown that these models can relate spoken words and sounds to visual content after training on a large-scale dataset of videos, but they were only trained and evaluated on videos in English. To learn multilingual audio-visual representations, we propose a cascaded approach that leverages a model trained on English videos and applies it to audio-visual data in other languages, such as Japanese videos. With our cascaded approach, we show an improvement in retrieval performance of nearly 10x compared to training on the Japanese videos solely. We also apply the model trained on English videos to Japanese and Hindi spoken captions of images, achieving state-of-the-art performance.

Puyuan Peng, David Harwath

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.

Changan Chen, Wei Sun, David Harwath et al.

Reverberation not only degrades the quality of speech for human perception, but also severely impacts the accuracy of automatic speech recognition. Prior work attempts to remove reverberation based on the audio modality only. Our idea is to learn to dereverberate speech from audio-visual observations. The visual environment surrounding a human speaker reveals important cues about the room geometry, materials, and speaker location, all of which influence the precise reverberation effects. We introduce Visually-Informed Dereverberation of Audio (VIDA), an end-to-end approach that learns to remove reverberation based on both the observed monaural sound and visual scene. In support of this new task, we develop a large-scale dataset SoundSpaces-Speech that uses realistic acoustic renderings of speech in real-world 3D scans of homes offering a variety of room acoustics. Demonstrating our approach on both simulated and real imagery for speech enhancement, speech recognition, and speaker identification, we show it achieves state-of-the-art performance and substantially improves over audio-only methods.

Mathew Monfort, SouYoung Jin, Alexander Liu et al.

When people observe events, they are able to abstract key information and build concise summaries of what is happening. These summaries include contextual and semantic information describing the important high-level details (what, where, who and how) of the observed event and exclude background information that is deemed unimportant to the observer. With this in mind, the descriptions people generate for videos of different dynamic events can greatly improve our understanding of the key information of interest in each video. These descriptions can be captured in captions that provide expanded attributes for video labeling (e.g. actions/objects/scenes/sentiment/etc.) while allowing us to gain new insight into what people find important or necessary to summarize specific events. Existing caption datasets for video understanding are either small in scale or restricted to a specific domain. To address this, we present the Spoken Moments (S-MiT) dataset of 500k spoken captions each attributed to a unique short video depicting a broad range of different events. We collect our descriptions using audio recordings to ensure that they remain as natural and concise as possible while allowing us to scale the size of a large classification dataset. In order to utilize our proposed dataset, we present a novel Adaptive Mean Margin (AMM) approach to contrastive learning and evaluate our models on video/caption retrieval on multiple datasets. We show that our AMM approach consistently improves our results and that models trained on our Spoken Moments dataset generalize better than those trained on other video-caption datasets.