Daily Papers

Full-duplex dialog models are designed to listen and speak simultaneously with rapid responses to fast-changing user input. Among existing approaches, native full-duplex models merges different channels (e.g. listen and speak) in a single time step, overcoming the high response latency inherent to time-division multiplexing time-division multiplexing (TDM) alternatives. Yet, a key challenge remains: aligning textual monologues with audio streams that operate at different bitrates. The prevailing solution relies on word-level alignment, but this can degrade the language ability of large pre-trained models. Moreover, it requires highly accurate timestamps for every token, which introduces cascading errors and increases pre-processing costs. In this paper, we propose textual monologues in continuous tokens sequence, namely "natural" monologues, which mimics humanoid cognitive behavior in dialogs. For temporal alignment, we alternate the position of the natural monologue - leading or trailing the audio - across different training stages. This "dual" training paradigm proves highly effective in building FLM-Audio, our 7B spoken dialog model that demonstrates superior responsiveness, duplexity, and chatting experiences, as confirmed by experimental results.

We introduce Delayed Streams Modeling (DSM), a flexible formulation for streaming, multimodal sequence-to-sequence learning. Sequence-to-sequence generation is often cast in an offline manner, where the model consumes the complete input sequence before generating the first output timestep. Alternatively, streaming sequence-to-sequence rely on learning a policy for choosing when to advance on the input stream, or write to the output stream. DSM instead models already time-aligned streams with a decoder-only language model. By moving the alignment to a pre-processing step,and introducing appropriate delays between streams, DSM provides streaming inference of arbitrary output sequences, from any input combination, making it applicable to many sequence-to-sequence problems. In particular, given text and audio streams, automatic speech recognition (ASR) corresponds to the text stream being delayed, while the opposite gives a text-to-speech (TTS) model. We perform extensive experiments for these two major sequence-to-sequence tasks, showing that DSM provides state-of-the-art performance and latency while supporting arbitrary long sequences, being even competitive with offline baselines. Code, samples and demos are available at https://github.com/kyutai-labs/delayed-streams-modeling

Streaming speech-to-text translation (StreamST) is the task of automatically translating speech while incrementally receiving an audio stream. Unlike simultaneous ST (SimulST), which deals with pre-segmented speech, StreamST faces the challenges of handling continuous and unbounded audio streams. This requires additional decisions about what to retain of the previous history, which is impractical to keep entirely due to latency and computational constraints. Despite the real-world demand for real-time ST, research on streaming translation remains limited, with existing works solely focusing on SimulST. To fill this gap, we introduce StreamAtt, the first StreamST policy, and propose StreamLAAL, the first StreamST latency metric designed to be comparable with existing metrics for SimulST. Extensive experiments across all 8 languages of MuST-C v1.0 show the effectiveness of StreamAtt compared to a naive streaming baseline and the related state-of-the-art SimulST policy, providing a first step in StreamST research.

We describe a system that generates speaker-annotated transcripts of meetings by using a virtual microphone array, a set of spatially distributed asynchronous recording devices such as laptops and mobile phones. The system is composed of continuous audio stream alignment, blind beamforming, speech recognition, speaker diarization using prior speaker information, and system combination. When utilizing seven input audio streams, our system achieves a word error rate (WER) of 22.3% and comes within 3% of the close-talking microphone WER on the non-overlapping speech segments. The speaker-attributed WER (SAWER) is 26.7%. The relative gains in SAWER over the single-device system are 14.8%, 20.3%, and 22.4% for three, five, and seven microphones, respectively. The presented system achieves a 13.6% diarization error rate when 10% of the speech duration contains more than one speaker. The contribution of each component to the overall performance is also investigated, and we validate the system with experiments on the NIST RT-07 conference meeting test set.

We present Beamforming-LLM, a system that enables users to semantically recall conversations they may have missed in multi-speaker environments. The system combines spatial audio capture using a microphone array with retrieval-augmented generation (RAG) to support natural language queries such as, "What did I miss when I was following the conversation on dogs?" Directional audio streams are separated using beamforming, transcribed with Whisper, and embedded into a vector database using sentence encoders. Upon receiving a user query, semantically relevant segments are retrieved, temporally aligned with non-attended segments, and summarized using a lightweight large language model (GPT-4o-mini). The result is a user-friendly interface that provides contrastive summaries, spatial context, and timestamped audio playback. This work lays the foundation for intelligent auditory memory systems and has broad applications in assistive technology, meeting summarization, and context-aware personal spatial computing.

Streaming Speech-to-Text Translation (StreamST) requires producing translations concurrently with incoming speech, imposing strict latency constraints and demanding models that balance partial-information decision-making with high translation quality. Research efforts on the topic have so far relied on the SimulEval repository, which is no longer maintained and does not support systems that revise their outputs. In addition, it has been designed for simulating the processing of short segments, rather than long-form audio streams, and it does not provide an easy method to showcase systems in a demo. As a solution, we introduce simulstream, the first open-source framework dedicated to unified evaluation and demonstration of StreamST systems. Designed for long-form speech processing, it supports not only incremental decoding approaches, but also re-translation methods, enabling for their comparison within the same framework both in terms of quality and latency. In addition, it also offers an interactive web interface to demo any system built within the tool.

There have been significant advances in deep learning for music demixing in recent years. However, there has been little attention given to how these neural networks can be adapted for real-time low-latency applications, which could be helpful for hearing aids, remixing audio streams and live shows. In this paper, we investigate the various challenges involved in adapting current demixing models in the literature for this use case. Subsequently, inspired by the Hybrid Demucs architecture, we propose the Hybrid Spectrogram Time-domain Audio Separation Network HS-TasNet, which utilises the advantages of spectral and waveform domains. For a latency of 23 ms, the HS-TasNet obtains an overall signal-to-distortion ratio (SDR) of 4.65 on the MusDB test set, and increases to 5.55 with additional training data. These results demonstrate the potential of efficient demixing for real-time low-latency music applications.

We consider the question: what can be learnt by looking at and listening to a large number of unlabelled videos? There is a valuable, but so far untapped, source of information contained in the video itself -- the correspondence between the visual and the audio streams, and we introduce a novel "Audio-Visual Correspondence" learning task that makes use of this. Training visual and audio networks from scratch, without any additional supervision other than the raw unconstrained videos themselves, is shown to successfully solve this task, and, more interestingly, result in good visual and audio representations. These features set the new state-of-the-art on two sound classification benchmarks, and perform on par with the state-of-the-art self-supervised approaches on ImageNet classification. We also demonstrate that the network is able to localize objects in both modalities, as well as perform fine-grained recognition tasks.

Radio remains a pervasive medium for mass information dissemination, with AM/FM stations reaching more Americans than either smartphone-based social networking or live television. Increasingly, radio broadcasts are also streamed online and accessed over the Internet. We present WavePulse, a framework that records, documents, and analyzes radio content in real-time. While our framework is generally applicable, we showcase the efficacy of WavePulse in a collaborative project with a team of political scientists focusing on the 2024 Presidential Elections. We use WavePulse to monitor livestreams of 396 news radio stations over a period of three months, processing close to 500,000 hours of audio streams. These streams were converted into time-stamped, diarized transcripts and analyzed to track answer key political science questions at both the national and state levels. Our analysis revealed how local issues interacted with national trends, providing insights into information flow. Our results demonstrate WavePulse's efficacy in capturing and analyzing content from radio livestreams sourced from the Web. Code and dataset can be accessed at https://wave-pulse.io.

Recent directions in automatic speech recognition (ASR) research have shown that applying deep learning models from image recognition challenges in computer vision is beneficial. As automatic music transcription (AMT) is superficially similar to ASR, in the sense that methods often rely on transforming spectrograms to symbolic sequences of events (e.g. words or notes), deep learning should benefit AMT as well. In this work, we outline an online polyphonic pitch detection system that streams audio to MIDI by ConvLSTMs. Our system achieves state-of-the-art results on the 2007 MIREX multi-F0 development set, with an F-measure of 83\% on the bassoon, clarinet, flute, horn and oboe ensemble recording without requiring any musical language modelling or assumptions of instrument timbre.

This paper focuses on two key problems for audio-visual emotion recognition in the video. One is the audio and visual streams temporal alignment for feature level fusion. The other one is locating and re-weighting the perception attentions in the whole audio-visual stream for better recognition. The Long Short Term Memory Recurrent Neural Network (LSTM-RNN) is employed as the main classification architecture. Firstly, soft attention mechanism aligns the audio and visual streams. Secondly, seven emotion embedding vectors, which are corresponding to each classification emotion type, are added to locate the perception attentions. The locating and re-weighting process is also based on the soft attention mechanism. The experiment results on EmotiW2015 dataset and the qualitative analysis show the efficiency of the proposed two techniques.

We introduce MAGNeT, a masked generative sequence modeling method that operates directly over several streams of audio tokens. Unlike prior work, MAGNeT is comprised of a single-stage, non-autoregressive transformer. During training, we predict spans of masked tokens obtained from a masking scheduler, while during inference we gradually construct the output sequence using several decoding steps. To further enhance the quality of the generated audio, we introduce a novel rescoring method in which, we leverage an external pre-trained model to rescore and rank predictions from MAGNeT, which will be then used for later decoding steps. Lastly, we explore a hybrid version of MAGNeT, in which we fuse between autoregressive and non-autoregressive models to generate the first few seconds in an autoregressive manner while the rest of the sequence is being decoded in parallel. We demonstrate the efficiency of MAGNeT for the task of text-to-music and text-to-audio generation and conduct an extensive empirical evaluation, considering both objective metrics and human studies. The proposed approach is comparable to the evaluated baselines, while being significantly faster (x7 faster than the autoregressive baseline). Through ablation studies and analysis, we shed light on the importance of each of the components comprising MAGNeT, together with pointing to the trade-offs between autoregressive and non-autoregressive modeling, considering latency, throughput, and generation quality. Samples are available on our demo page https://pages.cs.huji.ac.il/adiyoss-lab/MAGNeT.

Multi-media communications facilitate global interaction among people. However, despite researchers exploring cross-lingual translation techniques such as machine translation and audio speech translation to overcome language barriers, there is still a shortage of cross-lingual studies on visual speech. This lack of research is mainly due to the absence of datasets containing visual speech and translated text pairs. In this paper, we present AVMuST-TED, the first dataset for Audio-Visual Multilingual Speech Translation, derived from TED talks. Nonetheless, visual speech is not as distinguishable as audio speech, making it difficult to develop a mapping from source speech phonemes to the target language text. To address this issue, we propose MixSpeech, a cross-modality self-learning framework that utilizes audio speech to regularize the training of visual speech tasks. To further minimize the cross-modality gap and its impact on knowledge transfer, we suggest adopting mixed speech, which is created by interpolating audio and visual streams, along with a curriculum learning strategy to adjust the mixing ratio as needed. MixSpeech enhances speech translation in noisy environments, improving BLEU scores for four languages on AVMuST-TED by +1.4 to +4.2. Moreover, it achieves state-of-the-art performance in lip reading on CMLR (11.1\%), LRS2 (25.5\%), and LRS3 (28.0\%).

Multimodal question answering (QA) often requires identifying which video, audio, or sensor tokens are relevant to the question. Yet modality disagreements are common: off-camera speech, background noise, or motion outside the field of view often mislead fusion models that weight all streams equally. We present RAVEN, a unified QA architecture whose core is QuART, a query-conditioned cross-modal gating module that assigns scalar relevance scores to each token across modalities, enabling the model to amplify informative signals and suppress distractors before fusion. RAVEN is trained through a three-stage pipeline comprising unimodal pretraining, query-aligned fusion, and disagreement-oriented fine-tuning -- each stage targeting a distinct challenge in multi-modal reasoning: representation quality, cross-modal relevance, and robustness to modality mismatch. To support training and evaluation, we release AVS-QA, a dataset of 300K synchronized Audio--Video-Sensor streams paired with automatically generated question-answer pairs. Experimental results on seven multi-modal QA benchmarks -- including egocentric and exocentric tasks -- show that RAVEN achieves up to 14.5\% and 8.0\% gains in accuracy compared to state-of-the-art multi-modal large language models, respectively. Incorporating sensor data provides an additional 16.4\% boost, and the model remains robust under modality corruption, outperforming SOTA baselines by 50.23\%. Our code and dataset are available at https://github.com/BASHLab/RAVEN.

Portrait video editing focuses on modifying specific attributes of portrait videos, guided by audio or video streams. Previous methods typically either concentrate on lip-region reenactment or require training specialized models to extract keypoints for motion transfer to a new identity. In this paper, we introduce a training-free universal portrait video editing framework that provides a versatile and adaptable editing strategy. This framework supports portrait appearance editing conditioned on the changed first reference frame, as well as lip editing conditioned on varied speech, or a combination of both. It is based on a Unified Animation Control (UAC) mechanism with source inversion latents to edit the entire portrait, including visual-driven shape control, audio-driven speaking control, and inter-frame temporal control. Furthermore, our method can be adapted to different scenarios by adjusting the initial reference frame, enabling detailed editing of portrait videos with specific head rotations and facial expressions. This comprehensive approach ensures a holistic and flexible solution for portrait video editing. The experimental results show that our model can achieve more accurate and synchronized lip movements for the lip editing task, as well as more flexible motion transfer for the appearance editing task. Demo is available at https://alice01010101.github.io/RASA/.

Although audio-visual representation has been proved to be applicable in many downstream tasks, the representation of dancing videos, which is more specific and always accompanied by music with complex auditory contents, remains challenging and uninvestigated. Considering the intrinsic alignment between the cadent movement of dancer and music rhythm, we introduce MuDaR, a novel Music-Dance Representation learning framework to perform the synchronization of music and dance rhythms both in explicit and implicit ways. Specifically, we derive the dance rhythms based on visual appearance and motion cues inspired by the music rhythm analysis. Then the visual rhythms are temporally aligned with the music counterparts, which are extracted by the amplitude of sound intensity. Meanwhile, we exploit the implicit coherence of rhythms implied in audio and visual streams by contrastive learning. The model learns the joint embedding by predicting the temporal consistency between audio-visual pairs. The music-dance representation, together with the capability of detecting audio and visual rhythms, can further be applied to three downstream tasks: (a) dance classification, (b) music-dance retrieval, and (c) music-dance retargeting. Extensive experiments demonstrate that our proposed framework outperforms other self-supervised methods by a large margin.

Large Language Models (LLMs) have made significant strides in text generation and comprehension, with recent advancements extending into multimodal LLMs that integrate visual and audio inputs. However, these models continue to struggle with fine-grained, cross-modal temporal understanding, particularly when correlating events across audio and video streams. We address these challenges with two key contributions: a new dataset and model, called OCTAV and OMCAT respectively. OCTAV (Omni Context and Temporal Audio Video) is a novel dataset designed to capture event transitions across audio and video. Second, OMCAT (Omni Context Aware Transformer) is a powerful model that leverages RoTE (Rotary Time Embeddings), an innovative extension of RoPE, to enhance temporal grounding and computational efficiency in time-anchored tasks. Through a robust three-stage training pipeline-feature alignment, instruction tuning, and OCTAV-specific training-OMCAT excels in cross-modal temporal understanding. Our model demonstrates state-of-the-art performance on Audio-Visual Question Answering (AVQA) tasks and the OCTAV benchmark, showcasing significant gains in temporal reasoning and cross-modal alignment, as validated through comprehensive experiments and ablation studies. Our dataset and code will be made publicly available. The link to our demo page is https://om-cat.github.io.

Even though they differ in the physical domain, digital video and audio share many characteristics. Both are temporal data streams often stored in buffers with 8-bit values. This paper investigates a method for creating harmonic sounds with a video signal as input. A musical instrument is proposed, that utilizes video in both a sound synthesis method, and in a controller interface for selecting musical notes at specific velocities. The resulting instrument was informally determined by the author to sound both pleasant and interesting, but hard to control, and therefore suited for synth pad sounds.

Audio-visual large language models (LLM) have drawn significant attention, yet the fine-grained combination of both input streams is rather under-explored, which is challenging but necessary for LLMs to understand general video inputs. To this end, a fine-grained audio-visual joint representation (FAVOR) learning framework for multimodal LLMs is proposed in this paper, which extends a text-based LLM to simultaneously perceive speech and audio events in the audio input stream and images or videos in the visual input stream, at the frame level. To fuse the audio and visual feature streams into joint representations and to align the joint space with the LLM input embedding space, we propose a causal Q-Former structure with a causal attention module to enhance the capture of causal relations of the audio-visual frames across time. An audio-visual evaluation benchmark (AVEB) is also proposed which comprises six representative single-modal tasks with five cross-modal tasks reflecting audio-visual co-reasoning abilities. While achieving competitive single-modal performance on audio, speech and image tasks in AVEB, FAVOR achieved over 20% accuracy improvements on the video question-answering task when fine-grained information or temporal causal reasoning is required. FAVOR, in addition, demonstrated remarkable video comprehension and reasoning abilities on tasks that are unprecedented by other multimodal LLMs. An interactive demo of FAVOR is available at https://github.com/BriansIDP/AudioVisualLLM.git, and the training code and model checkpoints will be released soon.

Lack of audio-video synchronization is a common problem during television broadcasts and video conferencing, leading to an unsatisfactory viewing experience. A widely accepted paradigm is to create an error detection mechanism that identifies the cases when audio is leading or lagging. We propose ModEFormer, which independently extracts audio and video embeddings using modality-specific transformers. Different from the other transformer-based approaches, ModEFormer preserves the modality of the input streams which allows us to use a larger batch size with more negative audio samples for contrastive learning. Further, we propose a trade-off between the number of negative samples and number of unique samples in a batch to significantly exceed the performance of previous methods. Experimental results show that ModEFormer achieves state-of-the-art performance, 94.5% for LRS2 and 90.9% for LRS3. Finally, we demonstrate how ModEFormer can be used for offset detection for test clips.

Audio-visual event parsing plays a crucial role in understanding multimodal video content, but existing methods typically rely on offline processing of entire videos with huge model sizes, limiting their real-time applicability. We introduce Online Audio-Visual Event Parsing (On-AVEP), a novel paradigm for parsing audio, visual, and audio-visual events by sequentially analyzing incoming video streams. The On-AVEP task necessitates models with two key capabilities: (1) Accurate online inference, to effectively distinguish events with unclear and limited context in online settings, and (2) Real-time efficiency, to balance high performance with computational constraints. To cultivate these, we propose the Predictive Future Modeling (PreFM) framework featured by (a) predictive multimodal future modeling to infer and integrate beneficial future audio-visual cues, thereby enhancing contextual understanding and (b) modality-agnostic robust representation along with focal temporal prioritization to improve precision and generalization. Extensive experiments on the UnAV-100 and LLP datasets show PreFM significantly outperforms state-of-the-art methods by a large margin with significantly fewer parameters, offering an insightful approach for real-time multimodal video understanding. Code is available at https://github.com/XiaoYu-1123/PreFM.

In this paper, we present TalkingMachines -- an efficient framework that transforms pretrained video generation models into real-time, audio-driven character animators. TalkingMachines enables natural conversational experiences by integrating an audio large language model (LLM) with our video generation foundation model. Our primary contributions include: (1) We adapt a pretrained SOTA image-to-video DiT into an audio-driven avatar generation model of 18 billion parameters; (2) We enable infinite video streaming without error accumulation through asymmetric knowledge distillation from a bidirectional teacher model into a sparse causal, autoregressive student model; (3) We design a high-throughput, low-latency inference pipeline incorporating several key engineering optimizations such as: (a) disaggregation of the DiT and VAE decoder across separate devices, (b) efficient overlap of inter-device communication and computation using CUDA streams, (c) elimination of redundant recomputations to maximize frame-generation throughput. Please see demo videos here - https://aaxwaz.github.io/TalkingMachines/

Existing zero-shot text-to-speech (TTS) systems are typically designed to process complete sentences and are constrained by the maximum duration for which they have been trained. However, in many streaming applications, texts arrive continuously in short chunks, necessitating instant responses from the system. We identify the essential capabilities required for chunk-level streaming and introduce LiveSpeech 2, a stream-aware model that supports infinitely long speech generation, text-audio stream synchronization, and seamless transitions between short speech chunks. To achieve these, we propose (1) adopting Mamba, a class of sequence modeling distinguished by linear-time decoding, which is augmented by cross-attention mechanisms for conditioning, (2) utilizing rotary positional embeddings in the computation of cross-attention, enabling the model to process an infinite text stream by sliding a window, and (3) decoding with semantic guidance, a technique that aligns speech with the transcript during inference with minimal overhead. Experimental results demonstrate that our models are competitive with state-of-the-art language model-based zero-shot TTS models, while also providing flexibility to support a wide range of streaming scenarios.

With the emergence of neural audio codecs, which encode multiple streams of discrete tokens from audio, large language models have recently gained attention as a promising approach for zero-shot Text-to-Speech (TTS) synthesis. Despite the ongoing rush towards scaling paradigms, audio tokenization ironically amplifies the scalability challenge, stemming from its long sequence length and the complexity of modelling the multiple sequences. To mitigate these issues, we present CLaM-TTS that employs a probabilistic residual vector quantization to (1) achieve superior compression in the token length, and (2) allow a language model to generate multiple tokens at once, thereby eliminating the need for cascaded modeling to handle the number of token streams. Our experimental results demonstrate that CLaM-TTS is better than or comparable to state-of-the-art neural codec-based TTS models regarding naturalness, intelligibility, speaker similarity, and inference speed. In addition, we examine the impact of the pretraining extent of the language models and their text tokenization strategies on performances.

Automating the synthesis of coordinated bimanual piano performances poses significant challenges, particularly in capturing the intricate choreography between the hands while preserving their distinct kinematic signatures. In this paper, we propose a dual-stream neural framework designed to generate synchronized hand gestures for piano playing from audio input, addressing the critical challenge of modeling both hand independence and coordination. Our framework introduces two key innovations: (i) a decoupled diffusion-based generation framework that independently models each hand's motion via dual-noise initialization, sampling distinct latent noise for each while leveraging a shared positional condition, and (ii) a Hand-Coordinated Asymmetric Attention (HCAA) mechanism suppresses symmetric (common-mode) noise to highlight asymmetric hand-specific features, while adaptively enhancing inter-hand coordination during denoising. The system operates hierarchically: it first predicts 3D hand positions from audio features and then generates joint angles through position-aware diffusion models, where parallel denoising streams interact via HCAA. Comprehensive evaluations demonstrate that our framework outperforms existing state-of-the-art methods across multiple metrics.

The dynamic nature of esports makes the situation relatively complicated for average viewers. Esports broadcasting involves game expert casters, but the caster-dependent game commentary is not enough to fully understand the game situation. It will be richer by including diverse multimodal esports information, including audiences' talks/emotions, game audio, and game match event information. This paper introduces GAME-MUG, a new multimodal game situation understanding and audience-engaged commentary generation dataset and its strong baseline. Our dataset is collected from 2020-2022 LOL game live streams from YouTube and Twitch, and includes multimodal esports game information, including text, audio, and time-series event logs, for detecting the game situation. In addition, we also propose a new audience conversation augmented commentary dataset by covering the game situation and audience conversation understanding, and introducing a robust joint multimodal dual learning model as a baseline. We examine the model's game situation/event understanding ability and commentary generation capability to show the effectiveness of the multimodal aspects coverage and the joint integration learning approach.

We introduce SigmaCollab, a dataset enabling research on physically situated human-AI collaboration. The dataset consists of a set of 85 sessions in which untrained participants were guided by a mixed-reality assistive AI agent in performing procedural tasks in the physical world. SigmaCollab includes a set of rich, multimodal data streams, such as the participant and system audio, egocentric camera views from the head-mounted device, depth maps, head, hand and gaze tracking information, as well as additional annotations performed post-hoc. While the dataset is relatively small in size (~ 14 hours), its application-driven and interactive nature brings to the fore novel research challenges for human-AI collaboration, and provides more realistic testing grounds for various AI models operating in this space. In future work, we plan to use the dataset to construct a set of benchmarks for physically situated collaboration in mixed-reality task assistive scenarios. SigmaCollab is available at https://github.com/microsoft/SigmaCollab.

We propose Stable Video Infinity (SVI) that is able to generate infinite-length videos with high temporal consistency, plausible scene transitions, and controllable streaming storylines. While existing long-video methods attempt to mitigate accumulated errors via handcrafted anti-drifting (e.g., modified noise scheduler, frame anchoring), they remain limited to single-prompt extrapolation, producing homogeneous scenes with repetitive motions. We identify that the fundamental challenge extends beyond error accumulation to a critical discrepancy between the training assumption (seeing clean data) and the test-time autoregressive reality (conditioning on self-generated, error-prone outputs). To bridge this hypothesis gap, SVI incorporates Error-Recycling Fine-Tuning, a new type of efficient training that recycles the Diffusion Transformer (DiT)'s self-generated errors into supervisory prompts, thereby encouraging DiT to actively identify and correct its own errors. This is achieved by injecting, collecting, and banking errors through closed-loop recycling, autoregressively learning from error-injected feedback. Specifically, we (i) inject historical errors made by DiT to intervene on clean inputs, simulating error-accumulated trajectories in flow matching; (ii) efficiently approximate predictions with one-step bidirectional integration and calculate errors with residuals; (iii) dynamically bank errors into replay memory across discretized timesteps, which are resampled for new input. SVI is able to scale videos from seconds to infinite durations with no additional inference cost, while remaining compatible with diverse conditions (e.g., audio, skeleton, and text streams). We evaluate SVI on three benchmarks, including consistent, creative, and conditional settings, thoroughly verifying its versatility and state-of-the-art role.

We introduce Moshi, a speech-text foundation model and full-duplex spoken dialogue framework. Current systems for spoken dialogue rely on pipelines of independent components, namely voice activity detection, speech recognition, textual dialogue and text-to-speech. Such frameworks cannot emulate the experience of real conversations. First, their complexity induces a latency of several seconds between interactions. Second, text being the intermediate modality for dialogue, non-linguistic information that modifies meaning -- such as emotion or non-speech sounds -- is lost in the interaction. Finally, they rely on a segmentation into speaker turns, which does not take into account overlapping speech, interruptions and interjections. Moshi solves these independent issues altogether by casting spoken dialogue as speech-to-speech generation. Starting from a text language model backbone, Moshi generates speech as tokens from the residual quantizer of a neural audio codec, while modeling separately its own speech and that of the user into parallel streams. This allows for the removal of explicit speaker turns, and the modeling of arbitrary conversational dynamics. We moreover extend the hierarchical semantic-to-acoustic token generation of previous work to first predict time-aligned text tokens as a prefix to audio tokens. Not only this "Inner Monologue" method significantly improves the linguistic quality of generated speech, but we also illustrate how it can provide streaming speech recognition and text-to-speech. Our resulting model is the first real-time full-duplex spoken large language model, with a theoretical latency of 160ms, 200ms in practice, and is available at https://github.com/kyutai-labs/moshi.

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.

Recent text-to-video diffusion models can generate compelling video sequences, yet they remain silent -- missing the semantic, emotional, and atmospheric cues that audio provides. We introduce LTX-2, an open-source foundational model capable of generating high-quality, temporally synchronized audiovisual content in a unified manner. LTX-2 consists of an asymmetric dual-stream transformer with a 14B-parameter video stream and a 5B-parameter audio stream, coupled through bidirectional audio-video cross-attention layers with temporal positional embeddings and cross-modality AdaLN for shared timestep conditioning. This architecture enables efficient training and inference of a unified audiovisual model while allocating more capacity for video generation than audio generation. We employ a multilingual text encoder for broader prompt understanding and introduce a modality-aware classifier-free guidance (modality-CFG) mechanism for improved audiovisual alignment and controllability. Beyond generating speech, LTX-2 produces rich, coherent audio tracks that follow the characters, environment, style, and emotion of each scene -- complete with natural background and foley elements. In our evaluations, the model achieves state-of-the-art audiovisual quality and prompt adherence among open-source systems, while delivering results comparable to proprietary models at a fraction of their computational cost and inference time. All model weights and code are publicly released.

Automated Audio Captioning (AAC) is the task of generating natural language descriptions given an audio stream. A typical AAC system requires manually curated training data of audio segments and corresponding text caption annotations. The creation of these audio-caption pairs is costly, resulting in general data scarcity for the task. In this work, we address this major limitation and propose an approach to train AAC systems using only text. Our approach leverages the multimodal space of contrastively trained audio-text models, such as CLAP. During training, a decoder generates captions conditioned on the pretrained CLAP text encoder. During inference, the text encoder is replaced with the pretrained CLAP audio encoder. To bridge the modality gap between text and audio embeddings, we propose the use of noise injection or a learnable adapter, during training. We find that the proposed text-only framework performs competitively with state-of-the-art models trained with paired audio, showing that efficient text-to-audio transfer is possible. Finally, we showcase both stylized audio captioning and caption enrichment while training without audio or human-created text captions.

Speaker diarization, the process of segmenting an audio stream or transcribed speech content into homogenous partitions based on speaker identity, plays a crucial role in the interpretation and analysis of human speech. Most existing speaker diarization systems rely exclusively on unimodal acoustic information, making the task particularly challenging due to the innate ambiguities of audio signals. Recent studies have made tremendous efforts towards audio-visual or audio-semantic modeling to enhance performance. However, even the incorporation of up to two modalities often falls short in addressing the complexities of spontaneous and unstructured conversations. To exploit more meaningful dialogue patterns, we propose a novel multimodal approach that jointly utilizes audio, visual, and semantic cues to enhance speaker diarization. Our method elegantly formulates the multimodal modeling as a constrained optimization problem. First, we build insights into the visual connections among active speakers and the semantic interactions within spoken content, thereby establishing abundant pairwise constraints. Then we introduce a joint pairwise constraint propagation algorithm to cluster speakers based on these visual and semantic constraints. This integration effectively leverages the complementary strengths of different modalities, refining the affinity estimation between individual speaker embeddings. Extensive experiments conducted on multiple multimodal datasets demonstrate that our approach consistently outperforms state-of-the-art speaker diarization methods.

Recent advancements in video generation have seen a shift towards unified, transformer-based foundation models that can handle multiple conditional inputs in-context. However, these models have primarily focused on modalities like text, images, and depth maps, while strictly time-synchronous signals like audio have been underexplored. This paper introduces In-Context Audio Control of video diffusion transformers (ICAC), a framework that investigates the integration of audio signals for speech-driven video generation within a unified full-attention architecture, akin to FullDiT. We systematically explore three distinct mechanisms for injecting audio conditions: standard cross-attention, 2D self-attention, and unified 3D self-attention. Our findings reveal that while 3D attention offers the highest potential for capturing spatio-temporal audio-visual correlations, it presents significant training challenges. To overcome this, we propose a Masked 3D Attention mechanism that constrains the attention pattern to enforce temporal alignment, enabling stable training and superior performance. Our experiments demonstrate that this approach achieves strong lip synchronization and video quality, conditioned on an audio stream and reference images.

While automatic speech recognition (ASR) systems degrade significantly in noisy environments, audio-visual speech recognition (AVSR) systems aim to complement the audio stream with noise-invariant visual cues and improve the system's robustness. However, current studies mainly focus on fusing the well-learned modality features, like the output of modality-specific encoders, without considering the contextual relationship during the modality feature learning. In this study, we propose a multi-layer cross-attention fusion based AVSR (MLCA-AVSR) approach that promotes representation learning of each modality by fusing them at different levels of audio/visual encoders. Experimental results on the MISP2022-AVSR Challenge dataset show the efficacy of our proposed system, achieving a concatenated minimum permutation character error rate (cpCER) of 30.57% on the Eval set and yielding up to 3.17% relative improvement compared with our previous system which ranked the second place in the challenge. Following the fusion of multiple systems, our proposed approach surpasses the first-place system, establishing a new SOTA cpCER of 29.13% on this dataset.

Streaming speech recognition architectures are employed for low-latency, real-time applications. Such architectures are often characterized by their causality. Causal architectures emit tokens at each frame, relying only on current and past signal, while non-causal models are exposed to a window of future frames at each step to increase predictive accuracy. This dichotomy amounts to a trade-off for real-time Automatic Speech Recognition (ASR) system design: profit from the low-latency benefit of strictly-causal architectures while accepting predictive performance limitations, or realize the modeling benefits of future-context models accompanied by their higher latency penalty. In this work, we relax the constraints of this choice and present the Adaptive Non-Causal Attention Transducer (ANCAT). Our architecture is non-causal in the traditional sense, but executes in a low-latency, streaming manner by dynamically choosing when to rely on future context and to what degree within the audio stream. The resulting mechanism, when coupled with our novel regularization algorithms, delivers comparable accuracy to non-causal configurations while improving significantly upon latency, closing the gap with their causal counterparts. We showcase our design experimentally by reporting comparative ASR task results with measures of accuracy and latency on both publicly accessible and production-scale, voice-assistant datasets.

Audio-visual automatic speech recognition (AV-ASR) is an extension of ASR that incorporates visual cues, often from the movements of a speaker's mouth. Unlike works that simply focus on the lip motion, we investigate the contribution of entire visual frames (visual actions, objects, background etc.). This is particularly useful for unconstrained videos, where the speaker is not necessarily visible. To solve this task, we propose a new sequence-to-sequence AudioVisual ASR TrAnsformeR (AVATAR) which is trained end-to-end from spectrograms and full-frame RGB. To prevent the audio stream from dominating training, we propose different word-masking strategies, thereby encouraging our model to pay attention to the visual stream. We demonstrate the contribution of the visual modality on the How2 AV-ASR benchmark, especially in the presence of simulated noise, and show that our model outperforms all other prior work by a large margin. Finally, we also create a new, real-world test bed for AV-ASR called VisSpeech, which demonstrates the contribution of the visual modality under challenging audio conditions.

In this research paper, we delve into the topics of Speech Diarization and Automatic Speech Recognition (ASR). Speech diarization involves the separation of individual speakers within an audio stream. By employing the ASR transcript, the diarization process aims to segregate each speaker's utterances, grouping them based on their unique audio characteristics. On the other hand, Automatic Speech Recognition refers to the capability of a machine or program to identify and convert spoken words and phrases into a machine-readable format. In our speech diarization approach, we utilize the Gaussian Mixer Model (GMM) to represent speech segments. The inter-cluster distance is computed based on the GMM parameters, and the distance threshold serves as the stopping criterion. ASR entails the conversion of an unknown speech waveform into a corresponding written transcription. The speech signal is analyzed using synchronized algorithms, taking into account the pitch frequency. Our primary objective typically revolves around developing a model that minimizes the Word Error Rate (WER) metric during speech transcription.

Spoken keyword spotting (KWS) is the task of identifying a keyword in an audio stream and is widely used in smart devices at the edge in order to activate voice assistants and perform hands-free tasks. The task is daunting as there is a need, on the one hand, to achieve high accuracy while at the same time ensuring that such systems continue to run efficiently on low power and possibly limited computational capabilities devices. This work presents AraSpot for Arabic keyword spotting trained on 40 Arabic keywords, using different online data augmentation, and introducing ConformerGRU model architecture. Finally, we further improve the performance of the model by training a text-to-speech model for synthetic data generation. AraSpot achieved a State-of-the-Art SOTA 99.59% result outperforming previous approaches.

Voice activity detection is the task of detecting speech regions in a given audio stream or recording. First, we design a neural network combining trainable filters and recurrent layers to tackle voice activity detection directly from the waveform. Experiments on the challenging DIHARD dataset show that the proposed end-to-end model reaches state-of-the-art performance and outperforms a variant where trainable filters are replaced by standard cepstral coefficients. Our second contribution aims at making the proposed voice activity detection model robust to domain mismatch. To that end, a domain classification branch is added to the network and trained in an adversarial manner. The same DIHARD dataset, drawn from 11 different domains is used for evaluation under two scenarios. In the in-domain scenario where the training and test sets cover the exact same domains, we show that the domain-adversarial approach does not degrade performance of the proposed end-to-end model. In the out-domain scenario where the test domain is different from training domains, it brings a relative improvement of more than 10%. Finally, our last contribution is the provision of a fully reproducible open-source pipeline than can be easily adapted to other datasets.

Continuous speech separation for meeting pre-processing has recently become a focused research topic. Compared to the data in utterance-level speech separation, the meeting-style audio stream lasts longer, has an uncertain number of speakers. We adopt the time-domain speech separation method and the recently proposed Graph-PIT to build a super low-latency online speech separation model, which is very important for the real application. The low-latency time-domain encoder with a small stride leads to an extremely long feature sequence. We proposed a simple yet efficient model named Skipping Memory (SkiM) for the long sequence modeling. Experimental results show that SkiM achieves on par or even better separation performance than DPRNN. Meanwhile, the computational cost of SkiM is reduced by 75% compared to DPRNN. The strong long sequence modeling capability and low computational cost make SkiM a suitable model for online CSS applications. Our fastest real-time model gets 17.1 dB signal-to-distortion (SDR) improvement with less than 1-millisecond latency in the simulated meeting-style evaluation.

This paper explores the possibility of using visual object detection techniques for word localization in speech data. Object detection has been thoroughly studied in the contemporary literature for visual data. Noting that an audio can be interpreted as a 1-dimensional image, object localization techniques can be fundamentally useful for word localization. Building upon this idea, we propose a lightweight solution for word detection and localization. We use bounding box regression for word localization, which enables our model to detect the occurrence, offset, and duration of keywords in a given audio stream. We experiment with LibriSpeech and train a model to localize 1000 words. Compared to existing work, our method reduces model size by 94%, and improves the F1 score by 6.5\%.

Visual speech recognition (VSR) aims to recognize the content of speech based on lip movements, without relying on the audio stream. Advances in deep learning and the availability of large audio-visual datasets have led to the development of much more accurate and robust VSR models than ever before. However, these advances are usually due to the larger training sets rather than the model design. Here we demonstrate that designing better models is equally as important as using larger training sets. We propose the addition of prediction-based auxiliary tasks to a VSR model, and highlight the importance of hyperparameter optimization and appropriate data augmentations. We show that such a model works for different languages and outperforms all previous methods trained on publicly available datasets by a large margin. It even outperforms models that were trained on non-publicly available datasets containing up to to 21 times more data. We show, furthermore, that using additional training data, even in other languages or with automatically generated transcriptions, results in further improvement.

Textual overlays are often used in social media videos as people who watch them without the sound would otherwise miss essential information conveyed in the audio stream. This is why extraction of those overlays can serve as an important meta-data source, e.g. for content classification or retrieval tasks. In this work, we present a robust method for extracting textual overlays from videos that builds up on multiple neural network architectures. The proposed solution relies on several processing steps: keyframe extraction, text detection and text recognition. The main component of our system, i.e. the text recognition module, is inspired by a convolutional recurrent neural network architecture and we improve its performance using synthetically generated dataset of over 600,000 images with text prepared by authors specifically for this task. We also develop a filtering method that reduces the amount of overlapping text phrases using Levenshtein distance and further boosts system's performance. The final accuracy of our solution reaches over 80A% and is au pair with state-of-the-art methods.