Siddharth Gururani

Niket Agarwal, Arslan Ali, Maciej Bala et al. · nvidia

Physical AI needs to be trained digitally first. It needs a digital twin of itself, the policy model, and a digital twin of the world, the world model. In this paper, we present the Cosmos World Foundation Model Platform to help developers build customized world models for their Physical AI setups. We position a world foundation model as a general-purpose world model that can be fine-tuned into customized world models for downstream applications. Our platform covers a video curation pipeline, pre-trained world foundation models, examples of post-training of pre-trained world foundation models, and video tokenizers. To help Physical AI builders solve the most critical problems of our society, we make Cosmos open-source and our models open-weight with permissive licenses available via https://github.com/nvidia-cosmos/cosmos-predict1.

Arslan Ali, Junjie Bai, Maciej Bala et al. · nvidia

We introduce [Cosmos-Predict2.5], the latest generation of the Cosmos World Foundation Models for Physical AI. Built on a flow-based architecture, [Cosmos-Predict2.5] unifies Text2World, Image2World, and Video2World generation in a single model and leverages [Cosmos-Reason1], a Physical AI vision-language model, to provide richer text grounding and finer control of world simulation. Trained on 200M curated video clips and refined with reinforcement learning-based post-training, [Cosmos-Predict2.5] achieves substantial improvements over [Cosmos-Predict1] in video quality and instruction alignment, with models released at 2B and 14B scales. These capabilities enable more reliable synthetic data generation, policy evaluation, and closed-loop simulation for robotics and autonomous systems. We further extend the family with [Cosmos-Transfer2.5], a control-net style framework for Sim2Real and Real2Real world translation. Despite being 3.5$\times$ smaller than [Cosmos-Transfer1], it delivers higher fidelity and robust long-horizon video generation. Together, these advances establish [Cosmos-Predict2.5] and [Cosmos-Transfer2.5] as versatile tools for scaling embodied intelligence. To accelerate research and deployment in Physical AI, we release source code, pretrained checkpoints, and curated benchmarks under the NVIDIA Open Model License at https://github.com/nvidia-cosmos/cosmos-predict2.5 and https://github.com/nvidia-cosmos/cosmos-transfer2.5. We hope these open resources lower the barrier to adoption and foster innovation in building the next generation of embodied intelligence.

Sreyan Ghosh, Arushi Goel, Kaousheik Jayakumar et al.

We present Audio Flamingo Next (AF-Next), the next-generation and most capable large audio-language model in the Audio Flamingo series, designed to advance understanding and reasoning over speech, environmental sounds and music. Compared to Audio Flamingo 3, AF-Next introduces: (i) a stronger foundational audio-language model that significantly improves accuracy across diverse audio understanding tasks; (ii) scalable strategies for constructing large-scale audio understanding and reasoning data beyond existing academic benchmarks; (iii) support for long and complex audio inputs up to 30 minutes; and (iv) Temporal Audio Chain-of-Thought, a new reasoning paradigm that explicitly grounds intermediate reasoning steps to timestamps in long audio, enabling fine-grained temporal alignment and improved interpretability. To enable these capabilities, we first conduct a systematic analysis of Audio Flamingo 3 to identify key gaps in audio understanding and reasoning. We then curate and scale new large-scale datasets totaling over 1 million hours to address these limitations and expand the existing AudioSkills-XL, LongAudio-XL, AF-Think and AF-Chat datasets. AF-Next is trained using a curriculum-based strategy spanning pre-training, mid-training and post-training stages. Extensive experiments across 20 audio understanding and reasoning benchmarks, including challenging long-audio tasks, show that AF-Next outperforms similarly sized open models by large margins and remains highly competitive with and sometimes surpasses, much larger open-weight and closed models. Beyond benchmark performance, AF-Next exhibits strong real-world utility and transfers well to unseen tasks, highlighting its robustness and generalization ability. In addition to all data, code and methods, we open-source 3 variants of AF-Next, including AF-Next-Instruct, AF-Next-Think and AF-Next-Captioner.

Rohan Badlani, Rafael Valle, Kevin J. Shih et al.

We work to create a multilingual speech synthesis system which can generate speech with the proper accent while retaining the characteristics of an individual voice. This is challenging to do because it is expensive to obtain bilingual training data in multiple languages, and the lack of such data results in strong correlations that entangle speaker, language, and accent, resulting in poor transfer capabilities. To overcome this, we present a multilingual, multiaccented, multispeaker speech synthesis model based on RADTTS with explicit control over accent, language, speaker and fine-grained $F_0$ and energy features. Our proposed model does not rely on bilingual training data. We demonstrate an ability to control synthesized accent for any speaker in an open-source dataset comprising of 7 accents. Human subjective evaluation demonstrates that our model can better retain a speaker's voice and accent quality than controlled baselines while synthesizing fluent speech in all target languages and accents in our dataset.

Yuval Atzmon, Maciej Bala, Yogesh Balaji et al. · nvidia

We introduce Edify Image, a family of diffusion models capable of generating photorealistic image content with pixel-perfect accuracy. Edify Image utilizes cascaded pixel-space diffusion models trained using a novel Laplacian diffusion process, in which image signals at different frequency bands are attenuated at varying rates. Edify Image supports a wide range of applications, including text-to-image synthesis, 4K upsampling, ControlNets, 360 HDR panorama generation, and finetuning for image customization.

Arushi Goel, Sreyan Ghosh, Vatsal Agarwal et al.

Multimodal Large Language Models (MLLMs) have shown strong performance in visual and audio understanding when evaluated in isolation. However, their ability to jointly reason over omni-modal (visual, audio, and textual) signals in long and complex videos remains largely unexplored. We introduce MMOU, a new benchmark designed to systematically evaluate multimodal understanding and reasoning under these challenging, real-world conditions. MMOU consists of 15,000 carefully curated questions paired with 9038 web-collected videos of varying length, spanning diverse domains and exhibiting rich, tightly coupled audio-visual content. The benchmark covers 13 fundamental skill categories, all of which require integrating evidence across modalities and time. All questions are manually annotated across multiple turns by professional annotators, ensuring high quality and reasoning fidelity. We evaluate 20+ state-of-the-art open-source and proprietary multimodal models on MMOU. The results expose substantial performance gaps: the best closed-source model achieves only 64.2% accuracy, while the strongest open-source model reaches just 46.8%. Our results highlight the challenges of long-form omni-modal understanding, revealing that current models frequently fail to apply even fundamental skills in long videos. Through detailed analysis, we further identify systematic failure modes and provide insights into where and why current models break.

Siddharth Gururani, Arun Mallya, Ting-Chun Wang et al.

Animating portraits using speech has received growing attention in recent years, with various creative and practical use cases. An ideal generated video should have good lip sync with the audio, natural facial expressions and head motions, and high frame quality. In this work, we present SPACE, which uses speech and a single image to generate high-resolution, and expressive videos with realistic head pose, without requiring a driving video. It uses a multi-stage approach, combining the controllability of facial landmarks with the high-quality synthesis power of a pretrained face generator. SPACE also allows for the control of emotions and their intensities. Our method outperforms prior methods in objective metrics for image quality and facial motions and is strongly preferred by users in pair-wise comparisons. The project website is available at https://deepimagination.cc/SPACE/

Vinod Subramanian, Siddharth Gururani, Emmanouil Benetos et al.

Loss-gradients are used to interpret the decision making process of deep learning models. In this work, we evaluate loss-gradient based attribution methods by occluding parts of the input and comparing the performance of the occluded input to the original input. We observe that the occluded input has better performance than the original across the test dataset under certain conditions. Similar behaviour is observed in sound and image recognition tasks. We explore different loss-gradient attribution methods, occlusion levels and replacement values to explain the phenomenon of performance improvement under occlusion.

Yujia Huang, Adishree Ghatare, Yuanzhe Liu et al.

We study the problem of symbolic music generation (e.g., generating piano rolls), with a technical focus on non-differentiable rule guidance. Musical rules are often expressed in symbolic form on note characteristics, such as note density or chord progression, many of which are non-differentiable which pose a challenge when using them for guided diffusion. We propose Stochastic Control Guidance (SCG), a novel guidance method that only requires forward evaluation of rule functions that can work with pre-trained diffusion models in a plug-and-play way, thus achieving training-free guidance for non-differentiable rules for the first time. Additionally, we introduce a latent diffusion architecture for symbolic music generation with high time resolution, which can be composed with SCG in a plug-and-play fashion. Compared to standard strong baselines in symbolic music generation, this framework demonstrates marked advancements in music quality and rule-based controllability, outperforming current state-of-the-art generators in a variety of settings. For detailed demonstrations, code and model checkpoints, please visit our project website: https://scg-rule-guided-music.github.io/.

Siddharth Gururani, Alexander Lerch

Audio classification has seen great progress with the increasing availability of large-scale datasets. These large datasets, however, are often only partially labeled as collecting full annotations is a tedious and expensive process. This paper presents two semi-supervised methods capable of learning with missing labels and evaluates them on two publicly available, partially labeled datasets. The first method relies on label enhancement by a two-stage teacher-student learning process, while the second method utilizes the mean teacher semi-supervised learning algorithm. Our results demonstrate the impact of improperly handling missing labels and compare the benefits of using different strategies leveraging data with few labels. Methods capable of learning with partially labeled data have the potential to improve models for audio classification by utilizing even larger amounts of data without the need for complete annotations.

Jiawen Huang, Yun-Ning Hung, Ashis Pati et al.

The assessment of music performances in most cases takes into account the underlying musical score being performed. While there have been several automatic approaches for objective music performance assessment (MPA) based on extracted features from both the performance audio and the score, deep neural network-based methods incorporating score information into MPA models have not yet been investigated. In this paper, we introduce three different models capable of score-informed performance assessment. These are (i) a convolutional neural network that utilizes a simple time-series input comprising of aligned pitch contours and score, (ii) a joint embedding model which learns a joint latent space for pitch contours and scores, and (iii) a distance matrix-based convolutional neural network which utilizes patterns in the distance matrix between pitch contours and musical score to predict assessment ratings. Our results provide insights into the suitability of different architectures and input representations and demonstrate the benefits of score-informed models as compared to score-independent models.

Ashis Pati, Siddharth Gururani, Alexander Lerch

Representation learning focused on disentangling the underlying factors of variation in given data has become an important area of research in machine learning. However, most of the studies in this area have relied on datasets from the computer vision domain and thus, have not been readily extended to music. In this paper, we present a new symbolic music dataset that will help researchers working on disentanglement problems demonstrate the efficacy of their algorithms on diverse domains. This will also provide a means for evaluating algorithms specifically designed for music. To this end, we create a dataset comprising of 2-bar monophonic melodies where each melody is the result of a unique combination of nine latent factors that span ordinal, categorical, and binary types. The dataset is large enough (approx. 1.3 million data points) to train and test deep networks for disentanglement learning. In addition, we present benchmarking experiments using popular unsupervised disentanglement algorithms on this dataset and compare the results with those obtained on an image-based dataset.

Karn Watcharasupat, Siddharth Gururani, Alexander Lerch

In the field of music information retrieval, the task of simultaneously identifying the presence or absence of multiple musical instruments in a polyphonic recording remains a hard problem. Previous works have seen some success in improving instrument classification by applying temporal attention in a multi-instance multi-label setting, while another series of work has also suggested the role of pitch and timbre in improving instrument recognition performance. In this project, we further explore the use of attention mechanism in a timbral-temporal sense, à la visual attention, to improve the performance of musical instrument recognition using weakly-labeled data. Two approaches to this task have been explored. The first approach applies attention mechanism to the sliding-window paradigm, where a prediction based on each timbral-temporal `instance' is given an attention weight, before aggregation to produce the final prediction. The second approach is based on a recurrent model of visual attention where the network only attends to parts of the spectrogram and decide where to attend to next, given a limited number of `glimpses'.

Siddharth Gururani, Kilol Gupta, Dhaval Shah et al.

This paper presents a simple yet effective method to achieve prosody transfer from a reference speech signal to synthesized speech. The main idea is to incorporate well-known acoustic correlates of prosody such as pitch and loudness contours of the reference speech into a modern neural text-to-speech (TTS) synthesizer such as Tacotron2 (TC2). More specifically, a small set of acoustic features are extracted from reference audio and then used to condition a TC2 synthesizer. The trained model is evaluated using subjective listening tests and a novel objective evaluation of prosody transfer is proposed. Listening tests show that the synthesized speech is rated as highly natural and that prosody is successfully transferred from the reference speech signal to the synthesized signal.

Siddharth Gururani, Mohit Sharma, Alexander Lerch

While the automatic recognition of musical instruments has seen significant progress, the task is still considered hard for music featuring multiple instruments as opposed to single instrument recordings. Datasets for polyphonic instrument recognition can be categorized into roughly two categories. Some, such as MedleyDB, have strong per-frame instrument activity annotations but are usually small in size. Other, larger datasets such as OpenMIC only have weak labels, i.e., instrument presence or absence is annotated only for long snippets of a song. We explore an attention mechanism for handling weakly labeled data for multi-label instrument recognition. Attention has been found to perform well for other tasks with weakly labeled data. We compare the proposed attention model to multiple models which include a baseline binary relevance random forest, recurrent neural network, and fully connected neural networks. Our results show that incorporating attention leads to an overall improvement in classification accuracy metrics across all 20 instruments in the OpenMIC dataset. We find that attention enables models to focus on (or `attend to') specific time segments in the audio relevant to each instrument label leading to interpretable results.

Alexander Lerch, Claire Arthur, Ashis Pati et al.

Music Information Retrieval (MIR) tends to focus on the analysis of audio signals. Often, a single music recording is used as representative of a "song" even though different performances of the same song may reveal different properties. A performance is distinct in many ways from a (arguably more abstract) representation of a "song," "piece," or musical score. The characteristics of the (recorded) performance -- as opposed to the score or musical idea -- can have a major impact on how a listener perceives music. The analysis of music performance, however, has been traditionally only a peripheral topic for the MIR research community. This paper surveys the field of Music Performance Analysis (MPA) from various perspectives, discusses its significance to the field of MIR, and points out opportunities for future research in this field.