Search

Current emotional text-to-speech (TTS) systems face challenges in mimicking a broad spectrum of human emotions due to the inherent complexity of emotions and limitations in emotional speech datasets and models. This paper proposes a TTS framework that facilitates control over pleasure, arousal, and dominance, and can synthesize a diversity of emotional styles without requiring any emotional speech data during TTS training. We train an emotional attribute predictor using only categorical labels from speech data, aligning with psychological research and incorporating anchored dimensionality reduction on self-supervised learning (SSL) features. The TTS framework converts text inputs into phonetic tokens via an autoregressive language model and uses pseudo-emotional dimensions to guide the parallel prediction of fine-grained acoustic details. Experiments conducted on the LibriTTS dataset demonstrate that our framework can synthesize speech with enhanced naturalness and a variety of emotional styles by effectively controlling emotional dimensions, even without the inclusion of any emotional speech during TTS training., Comment: submitted to ICASSP 2025

Real-time cone-beam computed tomography (CBCT) provides instantaneous visualization of patient anatomy for image guidance, motion tracking, and online treatment adaptation in radiotherapy. While many real-time imaging and motion tracking methods leveraged patient-specific prior information to alleviate under-sampling challenges and meet the temporal constraint (< 500 ms), the prior information can be outdated and introduce biases, thus compromising the imaging and motion tracking accuracy. To address this challenge, we developed a framework (DREME) for real-time CBCT imaging and motion estimation, without relying on patient-specific prior knowledge. DREME incorporates a deep learning-based real-time CBCT imaging and motion estimation method into a dynamic CBCT reconstruction framework. The reconstruction framework reconstructs a dynamic sequence of CBCTs in a data-driven manner from a standard pre-treatment scan, without utilizing patient-specific knowledge. Meanwhile, a convolutional neural network-based motion encoder is jointly trained during the reconstruction to learn motion-related features relevant for real-time motion estimation, based on a single arbitrarily-angled x-ray projection. DREME was tested on digital phantom simulation and real patient studies. DREME accurately solved 3D respiration-induced anatomic motion in real time (~1.5 ms inference time for each x-ray projection). In the digital phantom study, it achieved an average lung tumor center-of-mass localization error of 1.2$\pm$0.9 mm (Mean$\pm$SD). In the patient study, it achieved a real-time tumor localization accuracy of 1.8$\pm$1.6 mm in the projection domain. DREME achieves CBCT and volumetric motion estimation in real time from a single x-ray projection at arbitrary angles, paving the way for future clinical applications in intra-fractional motion management.

With the advancements in singing voice generation and the growing presence of AI singers on media platforms, the inaugural Singing Voice Deepfake Detection (SVDD) Challenge aims to advance research in identifying AI-generated singing voices from authentic singers. This challenge features two tracks: a controlled setting track (CtrSVDD) and an in-the-wild scenario track (WildSVDD). The CtrSVDD track utilizes publicly available singing vocal data to generate deepfakes using state-of-the-art singing voice synthesis and conversion systems. Meanwhile, the WildSVDD track expands upon the existing SingFake dataset, which includes data sourced from popular user-generated content websites. For the CtrSVDD track, we received submissions from 47 teams, with 37 surpassing our baselines and the top team achieving a 1.65% equal error rate. For the WildSVDD track, we benchmarked the baselines. This paper reviews these results, discusses key findings, and outlines future directions for SVDD research., Comment: 6 pages, Accepted by 2024 IEEE Spoken Language Technology Workshop (SLT 2024)

This paper addresses the challenge of developing a robust audio-visual deepfake detection model. In practical use cases, new generation algorithms are continually emerging, and these algorithms are not encountered during the development of detection methods. This calls for the generalization ability of the method. Additionally, to ensure the credibility of detection methods, it is beneficial for the model to interpret which cues from the video indicate it is fake. Motivated by these considerations, we then propose a multi-stream fusion approach with one-class learning as a representation-level regularization technique. We study the generalization problem of audio-visual deepfake detection by creating a new benchmark by extending and re-splitting the existing FakeAVCeleb dataset. The benchmark contains four categories of fake videos (Real Audio-Fake Visual, Fake Audio-Fake Visual, Fake Audio-Real Visual, and Unsynchronized videos). The experimental results demonstrate that our approach surpasses the previous models by a large margin. Furthermore, our proposed framework offers interpretability, indicating which modality the model identifies as more likely to be fake. The source code is released at https://github.com/bok-bok/MSOC.

Recent singing voice synthesis and conversion advancements necessitate robust singing voice deepfake detection (SVDD) models. Current SVDD datasets face challenges due to limited controllability, diversity in deepfake methods, and licensing restrictions. Addressing these gaps, we introduce CtrSVDD, a large-scale, diverse collection of bonafide and deepfake singing vocals. These vocals are synthesized using state-of-the-art methods from publicly accessible singing voice datasets. CtrSVDD includes 47.64 hours of bonafide and 260.34 hours of deepfake singing vocals, spanning 14 deepfake methods and involving 164 singer identities. We also present a baseline system with flexible front-end features, evaluated against a structured train/dev/eval split. The experiments show the importance of feature selection and highlight a need for generalization towards deepfake methods that deviate further from training distribution. The CtrSVDD dataset and baselines are publicly accessible., Comment: Accepted by Interspeech 2024

The rapid advancement of AI-generated singing voices, which now closely mimic natural human singing and align seamlessly with musical scores, has led to heightened concerns for artists and the music industry. Unlike spoken voice, singing voice presents unique challenges due to its musical nature and the presence of strong background music, making singing voice deepfake detection (SVDD) a specialized field requiring focused attention. To promote SVDD research, we recently proposed the "SVDD Challenge," the very first research challenge focusing on SVDD for lab-controlled and in-the-wild bonafide and deepfake singing voice recordings. The challenge will be held in conjunction with the 2024 IEEE Spoken Language Technology Workshop (SLT 2024)., Comment: Evaluation plan of the SVDD Challenge @ SLT 2024

Cone-beam computed tomography (CBCT) is widely used in image-guided radiotherapy. Reconstructing CBCTs from limited-angle acquisitions (LA-CBCT) is highly desired for improved imaging efficiency, dose reduction, and better mechanical clearance. LA-CBCT reconstruction, however, suffers from severe under-sampling artifacts, making it a highly ill-posed inverse problem. Diffusion models can generate data/images by reversing a data-noising process through learned data distributions; and can be incorporated as a denoiser/regularizer in LA-CBCT reconstruction. In this study, we developed a diffusion model-based framework, prior frequency-guided diffusion model (PFGDM), for robust and structure-preserving LA-CBCT reconstruction. PFGDM uses a conditioned diffusion model as a regularizer for LA-CBCT reconstruction, and the condition is based on high-frequency information extracted from patient-specific prior CT scans which provides a strong anatomical prior for LA-CBCT reconstruction. Specifically, we developed two variants of PFGDM (PFGDM-A and PFGDM-B) with different conditioning schemes. PFGDM-A applies the high-frequency CT information condition until a pre-optimized iteration step, and drops it afterwards to enable both similar and differing CT/CBCT anatomies to be reconstructed. PFGDM-B, on the other hand, continuously applies the prior CT information condition in every reconstruction step, while with a decaying mechanism, to gradually phase out the reconstruction guidance from the prior CT scans. The two variants of PFGDM were tested and compared with current available LA-CBCT reconstruction solutions, via metrics including PSNR and SSIM. PFGDM outperformed all traditional and diffusion model-based methods. PFGDM reconstructs high-quality LA-CBCTs under very-limited gantry angles, allowing faster and more flexible CBCT scans with dose reductions., Comment: 20 pages, 8 figures, submitted to Physics in Medicine & Biology

Effectively and efficiently adapting a pre-trained language model (PLM) for human-centered text understanding (HCTU) is challenging since user tokens are million-level in most personalized applications and do not have concrete explicit semantics. A standard and parameter-efficient approach (e.g., LoRA) necessitates memorizing numerous suits of adapters for each user. In this work, we introduce a personalized LoRA (PLoRA) with a plug-and-play (PnP) framework for the HCTU task. PLoRA is effective, parameter-efficient, and dynamically deploying in PLMs. Moreover, a personalized dropout and a mutual information maximizing strategies are adopted and hence the proposed PLoRA can be well adapted to few/zero-shot learning scenarios for the cold-start issue. Experiments conducted on four benchmark datasets show that the proposed method outperforms existing methods in full/few/zero-shot learning scenarios for the HCTU task, even though it has fewer trainable parameters. For reproducibility, the code for this paper is available at: https://github.com/yoyo-yun/PLoRA., Comment: Accepted by AAAI 2024

Autonomous driving is an emerging technology that has advanced rapidly over the last decade. Modern transportation is expected to benefit greatly from a wise decision-making framework of autonomous vehicles, including the improvement of mobility and the minimization of risks and travel time. However, existing methods either ignore the complexity of environments only fitting straight roads, or ignore the impact on surrounding vehicles during optimization phases, leading to weak environmental adaptability and incomplete optimization objectives. To address these limitations, we propose a parameterized decision-making framework with multi-modal perception based on deep reinforcement learning, called AUTO. We conduct a comprehensive perception to capture the state features of various traffic participants around the autonomous vehicle, based on which we design a graph-based model to learn a state representation of the multi-modal semantic features. To distinguish between lane-following and lane-changing, we decompose an action of the autonomous vehicle into a parameterized action structure that first decides whether to change lanes and then computes an exact action to execute. A hybrid reward function takes into account aspects of safety, traffic efficiency, passenger comfort, and impact to guide the framework to generate optimal actions. In addition, we design a regularization term and a multi-worker paradigm to enhance the training. Extensive experiments offer evidence that AUTO can advance state-of-the-art in terms of both macroscopic and microscopic effectiveness., Comment: IEEE International Conference on Data Engineering (ICDE2024)

Dimensional representations of speech emotions such as the arousal-valence (AV) representation provide a continuous and fine-grained description and control than their categorical counterparts. They have wide applications in tasks such as dynamic emotion understanding and expressive text-to-speech synthesis. Existing methods that predict the dimensional emotion representation from speech cast it as a supervised regression task. These methods face data scarcity issues, as dimensional annotations are much harder to acquire than categorical labels. In this work, we propose to learn the AV representation from categorical emotion labels of speech. We start by learning a rich and emotion-relevant high-dimensional speech feature representation using self-supervised pre-training and emotion classification fine-tuning. This representation is then mapped to the 2D AV space according to psychological findings through anchored dimensionality reduction. Experiments show that our method achieves a Concordance Correlation Coefficient (CCC) performance comparable to state-of-the-art supervised regression methods on IEMOCAP without leveraging ground-truth AV annotations during training. This validates our proposed approach on AV prediction. Furthermore, visualization of AV predictions on MEAD and EmoDB datasets shows the interpretability of the learned AV representations.

Diffusion models have demonstrated significant potential in producing high-quality images in medical image translation to aid disease diagnosis, localization, and treatment. Nevertheless, current diffusion models have limited success in achieving faithful image translations that can accurately preserve the anatomical structures of medical images, especially for unpaired datasets. The preservation of structural and anatomical details is essential to reliable medical diagnosis and treatment planning, as structural mismatches can lead to disease misidentification and treatment errors. In this study, we introduce the Frequency Decoupled Diffusion Model (FDDM) for MR-to-CT conversion. FDDM first obtains the anatomical information of the CT image from the MR image through an initial conversion module. This anatomical information then guides a subsequent diffusion model to generate high-quality CT images. Our diffusion model uses a dual-path reverse diffusion process for low-frequency and high-frequency information, achieving a better balance between image quality and anatomical accuracy. We extensively evaluated FDDM using public datasets for brain MR-to-CT and pelvis MR-to-CT translations, demonstrating its superior performance to other GAN-based, VAE-based, and diffusion-based models. The evaluation metrics included Frechet Inception Distance (FID), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM). FDDM achieved the best scores on all metrics for both datasets, particularly excelling in FID, with scores of 25.9 for brain data and 29.2 for pelvis data, significantly outperforming other methods. These results demonstrate that FDDM can generate high-quality target domain images while maintaining the accuracy of translated anatomical structures.

Dynamic cone-beam computed tomography (CBCT) can capture high-spatial-resolution, time-varying images for motion monitoring, patient setup, and adaptive planning of radiotherapy. However, dynamic CBCT reconstruction is an extremely ill-posed spatiotemporal inverse problem, as each CBCT volume in the dynamic sequence is only captured by one or a few X-ray projections. We developed a machine learning-based technique, prior-model-free spatiotemporal implicit neural representation (PMF-STINR), to reconstruct dynamic CBCTs from sequentially acquired X-ray projections. PMF-STINR employs a joint image reconstruction and registration approach to address the under-sampling challenge. Specifically, PMF-STINR uses spatial implicit neural representation to reconstruct a reference CBCT volume, and it applies temporal INR to represent the intra-scan dynamic motion with respect to the reference CBCT to yield dynamic CBCTs. PMF-STINR couples the temporal INR with a learning-based B-spline motion model to capture time-varying deformable motion during the reconstruction. Compared with previous methods, the spatial INR, the temporal INR, and the B-spline model of PMF-STINR are all learned on the fly during reconstruction in a one-shot fashion, without using any patient-specific prior knowledge or motion sorting/binning. PMF-STINR was evaluated via digital phantom simulations, physical phantom measurements, and a multi-institutional patient dataset featuring various imaging protocols (half-fan/full-fan, full sampling/sparse sampling, different energy and mAs settings, etc.). The results showed that the one-shot learning-based PMF-STINR can accurately and robustly reconstruct dynamic CBCTs and capture highly irregular motion with high temporal (~0.1s) resolution and sub-millimeter accuracy. It can be a promising tool for motion management by offering richer motion information than traditional 4D-CBCTs.

Automatic segmentation of medical images is crucial in modern clinical workflows. The Segment Anything Model (SAM) has emerged as a versatile tool for image segmentation without specific domain training, but it requires human prompts and may have limitations in specific domains. Traditional models like nnUNet perform automatic segmentation during inference and are effective in specific domains but need extensive domain-specific training. To combine the strengths of foundational and domain-specific models, we propose nnSAM, integrating SAM's robust feature extraction with nnUNet's automatic configuration to enhance segmentation accuracy on small datasets. Our nnSAM model optimizes two main approaches: leveraging SAM's feature extraction and nnUNet's domain-specific adaptation, and incorporating a boundary shape supervision loss function based on level set functions and curvature calculations to learn anatomical shape priors from limited data. We evaluated nnSAM on four segmentation tasks: brain white matter, liver, lung, and heart segmentation. Our method outperformed others, achieving the highest DICE score of 82.77% and the lowest ASD of 1.14 mm in brain white matter segmentation with 20 training samples, compared to nnUNet's DICE score of 79.25% and ASD of 1.36 mm. A sample size study highlighted nnSAM's advantage with fewer training samples. Our results demonstrate significant improvements in segmentation performance with nnSAM, showcasing its potential for small-sample learning in medical image segmentation.

The rise of singing voice synthesis presents critical challenges to artists and industry stakeholders over unauthorized voice usage. Unlike synthesized speech, synthesized singing voices are typically released in songs containing strong background music that may hide synthesis artifacts. Additionally, singing voices present different acoustic and linguistic characteristics from speech utterances. These unique properties make singing voice deepfake detection a relevant but significantly different problem from synthetic speech detection. In this work, we propose the singing voice deepfake detection task. We first present SingFake, the first curated in-the-wild dataset consisting of 28.93 hours of bonafide and 29.40 hours of deepfake song clips in five languages from 40 singers. We provide a train/validation/test split where the test sets include various scenarios. We then use SingFake to evaluate four state-of-the-art speech countermeasure systems trained on speech utterances. We find these systems lag significantly behind their performance on speech test data. When trained on SingFake, either using separated vocal tracks or song mixtures, these systems show substantial improvement. However, our evaluations also identify challenges associated with unseen singers, communication codecs, languages, and musical contexts, calling for dedicated research into singing voice deepfake detection. The SingFake dataset and related resources are available at https://www.singfake.org/., Comment: Accepted at ICASSP 2024

The reconstruction of 3D cine-MRI is challenged by highly undersampled k-space data in each cine frame, due to the slow speed of MR signal acquisition. We proposed a machine learning-based framework, spatial and temporal implicit neural representation learning (STINR-MR), for accurate 3D cine-MRI reconstruction from highly undersampled data. STINR-MR used a joint reconstruction and deformable registration approach to address the ill-posed spatiotemporal reconstruction problem, by solving a reference-frame 3D MR image and a corresponding motion model which deforms the reference frame to each cine frame. The reference-frame image was reconstructed as a spatial implicit neural representation (INR) network, which learns the mapping from input 3D spatial coordinates to corresponding MR values. The dynamic motion model was constructed via a temporal INR, as well as basis deformation vector fields(DVFs) extracted from prior/onboard 4D-MRIs. The learned INR encodes input time points and outputs corresponding weighting factors to combine the basis DVFs into time-resolved motion fields. STINR-MR was evaluated using MR data simulated from the 4D extended cardiac-torso (XCAT) digital phantom and MR data acquired clinically from a healthy human subject. Its reconstruction accuracy was also compared with that of the model-based non-rigid motion estimation method (MR-MOTUS). STINR-MR can reconstruct 3D cine-MR images with high temporal (<100 ms) and spatial (3 mm) resolutions to accurately capture different irregular motion patterns. Compared with MR-MOTUS, STINR-MR consistently reconstructed images with better quality, fewer artifacts, and achieved superior tumor localization accuracy. STINR-MR provides a lightweight and efficient framework for accurate 3D cine-MRI reconstruction, and does not require external data for pre-training to avoid generalizability issues encountered in deep learning methods.

Individualized head-related transfer functions (HRTFs) are crucial for accurate sound positioning in virtual auditory displays. As the acoustic measurement of HRTFs is resource-intensive, predicting individualized HRTFs using machine learning models is a promising approach at scale. Training such models require a unified HRTF representation across multiple databases to utilize their respectively limited samples. However, in addition to differences on the spatial sampling locations, recent studies have shown that, even for the common location, HRTFs across databases manifest consistent differences that make it trivial to tell which databases they come from. This poses a significant challenge for learning a unified HRTF representation across databases. In this work, we first identify the possible causes of these cross-database differences, attributing them to variations in the measurement setup. Then, we propose a novel approach to normalize the frequency responses of HRTFs across databases. We show that HRTFs from different databases cannot be classified by their database after normalization. We further show that these normalized HRTFs can be used to learn a more unified HRTF representation across databases than the prior art. We believe that this normalization approach paves the road to many data-intensive tasks on HRTF modeling., Comment: 5 pages, 4 figures, accepted by IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA) 2023

In this paper, we aim to address the problem of channel robustness in speech countermeasure (CM) systems, which are used to distinguish synthetic speech from human natural speech. On the basis of two hypotheses, we suggest an approach for perturbing phase information during the training of time-domain CM systems. Communication networks often employ lossy compression codec that encodes only magnitude information, therefore heavily altering phase information. Also, state-of-the-art CM systems rely on phase information to identify spoofed speech. Thus, we believe the information loss in the phase domain induced by lossy compression codec degrades the performance of the unseen channel. We first establish the dependence of time-domain CM systems on phase information by perturbing phase in evaluation, showing strong degradation. Then, we demonstrated that perturbing phase during training leads to a significant performance improvement, whereas perturbing magnitude leads to further degradation., Comment: 5 pages; Proceedings of Interspeech 2023

Image translation has wide applications, such as style transfer and modality conversion, usually aiming to generate images having both high degrees of realism and faithfulness. These problems remain difficult, especially when it is important to preserve semantic structures. Traditional image-level similarity metrics are of limited use, since the semantics of an image are high-level, and not strongly governed by pixel-wise faithfulness to an original image. Towards filling this gap, we introduce SAMScore, a generic semantic structural similarity metric for evaluating the faithfulness of image translation models. SAMScore is based on the recent high-performance Segment Anything Model (SAM), which can perform semantic similarity comparisons with standout accuracy. We applied SAMScore on 19 image translation tasks, and found that it is able to outperform all other competitive metrics on all of the tasks. We envision that SAMScore will prove to be a valuable tool that will help to drive the vibrant field of image translation, by allowing for more precise evaluations of new and evolving translation models. The code is available at https://github.com/Kent0n-Li/SAMScore.