Hanlin Yu

Ruinan Jin, Xinting Liao, Hanlin Yu et al.

Modern voice cloning, also known as zero-shot text-to-speech (TTS), can synthesize speech that closely matches a target speaker from only seconds of reference audio, enabling applications such as personalized speech interfaces and dubbing. In practice, these systems often face noisy reference audio, imperfect text prompts, multilingual and long-form generation, post-processing, and adversarial perturbations, all of which can weaken robustness. Despite rapid progress in codec-token language models and diffusion-based TTS, robustness under realistic deployment shifts remains underexplored. This paper introduces RVCBench, a comprehensive dataset and benchmark for evaluating robustness in voice cloning. RVCBench provides task-aligned tests covering controlled text-audio pairing, multilingual and long-form scenarios, expressive prompts, post-processing conditions, and passive or proactive audio perturbations. Across 18 robustness evaluations, 225 speakers, and 14,370 utterances, RVCBench supports unified evaluation of input sensitivity, generation stability, output resilience, perturbation robustness, speaker similarity, and deepfake detectability. We evaluate 18 representative open-source voice cloning models and reveal systematic vulnerabilities in content consistency, speaker similarity, long-form stability, post-processing resilience, adversarial robustness, and detector-facing separability. We release the code and dataset to support reproducible evaluation and future research on robust voice cloning, speech synthesis, and audio generation. Code: https://github.com/Nanboy-Ronan/RVCBench. Dataset: https://huggingface.co/datasets/Nanboy/RVCBench.

Marcelo Hartmann, Bernardo Williams, Hanlin Yu et al.

We consider the fundamental task of optimising a real-valued function defined in a potentially high-dimensional Euclidean space, such as the loss function in many machine-learning tasks or the logarithm of the probability distribution in statistical inference. We use Riemannian geometry notions to redefine the optimisation problem of a function on the Euclidean space to a Riemannian manifold with a warped metric, and then find the function's optimum along this manifold. The warped metric chosen for the search domain induces a computational friendly metric-tensor for which optimal search directions associated with geodesic curves on the manifold becomes easier to compute. Performing optimization along geodesics is known to be generally infeasible, yet we show that in this specific manifold we can analytically derive Taylor approximations up to third-order. In general these approximations to the geodesic curve will not lie on the manifold, however we construct suitable retraction maps to pull them back onto the manifold. Therefore, we can efficiently optimize along the approximate geodesic curves. We cover the related theory, describe a practical optimization algorithm and empirically evaluate it on a collection of challenging optimisation benchmarks. Our proposed algorithm, using 3rd-order approximation of geodesics, tends to outperform standard Euclidean gradient-based counterparts in term of number of iterations until convergence.

Hanlin Yu, Marcelo Hartmann, Bernardo Williams et al.

Stochastic-gradient sampling methods are often used to perform Bayesian inference on neural networks. It has been observed that the methods in which notions of differential geometry are included tend to have better performances, with the Riemannian metric improving posterior exploration by accounting for the local curvature. However, the existing methods often resort to simple diagonal metrics to remain computationally efficient. This loses some of the gains. We propose two non-diagonal metrics that can be used in stochastic-gradient samplers to improve convergence and exploration but have only a minor computational overhead over diagonal metrics. We show that for fully connected neural networks (NNs) with sparsity-inducing priors and convolutional NNs with correlated priors, using these metrics can provide improvements. For some other choices the posterior is sufficiently easy also for the simpler metrics.

Hanlin Yu, Marcelo Hartmann, Bernardo Williams et al.

Laplace's method approximates a target density with a Gaussian distribution at its mode. It is computationally efficient and asymptotically exact for Bayesian inference due to the Bernstein-von Mises theorem, but for complex targets and finite-data posteriors it is often too crude an approximation. A recent generalization of the Laplace Approximation transforms the Gaussian approximation according to a chosen Riemannian geometry providing a richer approximation family, while still retaining computational efficiency. However, as shown here, its properties depend heavily on the chosen metric, indeed the metric adopted in previous work results in approximations that are overly narrow as well as being biased even at the limit of infinite data. We correct this shortcoming by developing the approximation family further, deriving two alternative variants that are exact at the limit of infinite data, extending the theoretical analysis of the method, and demonstrating practical improvements in a range of experiments.

Hanlin Yu, RuiKang OuYang, Partha Kaushik et al.

Learning an energy-based model from data samples is a central problem in machine learning. Many recent and popular methods, such as denoising score matching for training energy-based diffusion models, use stochastic interpolants to corrupt data samples at different noise levels indexed by a time variable. This defines a joint density over both the data space and time, and most methods learn its energy through either spatial or temporal differences. We identify distinct failure modes for both of these approaches. To solve them, we propose Spatiotemporal Noise-Contrastive Estimation (stNCE), a framework for learning the energy through joint spatiotemporal differences. stNCE unifies many existing methods and leads to new training objectives. Experiments on images and molecules demonstrate performance competitive with state-of-the-art density estimation methods.

Hanlin Yu, Arto Klami, Aapo Hyvärinen et al.

Density ratio estimation in high dimensions can be reframed as integrating a certain quantity, the time score, over probability paths which interpolate between the two densities. In practice, the time score has to be estimated based on samples from the two densities. However, existing methods for this problem remain computationally expensive and can yield inaccurate estimates. Inspired by recent advances in generative modeling, we introduce a novel framework for time score estimation, based on a conditioning variable. Choosing the conditioning variable judiciously enables a closed-form objective function. We demonstrate that, compared to previous approaches, our approach results in faster learning of the time score and competitive or better estimation accuracies of the density ratio on challenging tasks. Furthermore, we establish theoretical guarantees on the error of the estimated density ratio.

Hanlin Yu, Berfin Inal, Georgios Arvanitidis et al.

Neural models learn representations of high-dimensional data on low-dimensional manifolds. Multiple factors, including stochasticities in the training process, model architectures, and additional inductive biases, may induce different representations, even when learning the same task on the same data. However, it has recently been shown that when a latent structure is shared between distinct latent spaces, relative distances between representations can be preserved, up to distortions. Building on this idea, we demonstrate that exploiting the differential-geometric structure of latent spaces of neural models, it is possible to capture precisely the transformations between representational spaces trained on similar data distributions. Specifically, we assume that distinct neural models parametrize approximately the same underlying manifold, and introduce a representation based on the pullback metric that captures the intrinsic structure of the latent space, while scaling efficiently to large models. We validate experimentally our method on model stitching and retrieval tasks, covering autoencoders and vision foundation discriminative models, across diverse architectures, datasets, and pretraining schemes.

Hanlin Yu, Søren Hauberg, Marcelo Hartmann et al.

Real world data often lie on low-dimensional Riemannian manifolds embedded in high-dimensional spaces. This motivates learning degenerate normalizing flows that map between the ambient space and a low-dimensional latent space. However, if the manifold has a non-trivial topology, it can never be correctly learned using a single flow. Instead multiple flows must be `glued together'. In this paper, we first propose the general training scheme for learning such a collection of flows, and secondly we develop the first numerical algorithms for computing geodesics on such manifolds. Empirically, we demonstrate that this leads to highly significant improvements in topology estimation.

Boxuan Cao, Linkai Li, Hanlin Yu et al.

Speech intelligibility evaluation for hearing-impaired (HI) listeners is essential for assessing hearing aid performance, traditionally relying on listening tests or intrusive methods like HASPI. However, these methods require clean reference signals, which are often unavailable in real-world conditions, creating a gap between lab-based and real-world assessments. To address this, we propose a non-intrusive intelligibility prediction framework that leverages speech enhancers to provide a parallel enhanced-signal pathway, enabling robust predictions without reference signals. We evaluate three state-of-the-art enhancers and demonstrate that prediction performance depends on the choice of enhancer, with ensembles of strong enhancers yielding the best results. To improve cross-dataset generalization, we introduce a 2-clips augmentation strategy that enhances listener-specific variability, boosting robustness on unseen datasets. Our approach consistently outperforms the non-intrusive baseline, CPC2 Champion across multiple datasets, highlighting the potential of enhancer-guided non-intrusive intelligibility prediction for real-world applications.

Bernardo Williams, Hanlin Yu, Hoang Phuc Hau Luu et al.

Traditional Markov Chain Monte Carlo sampling methods often struggle with sharp curvatures, intricate geometries, and multimodal distributions. Slice sampling can resolve local exploration inefficiency issues, and Riemannian geometries help with sharp curvatures. Recent extensions enable slice sampling on Riemannian manifolds, but they are restricted to cases where geodesics are available in a closed form. We propose a method that generalizes Hit-and-Run slice sampling to more general geometries tailored to the target distribution, by approximating geodesics as solutions to differential equations. Our approach enables the exploration of the regions with strong curvature and rapid transitions between modes in multimodal distributions. We demonstrate the advantages of the approach over challenging sampling problems.

Hoang Phuc Hau Luu, Hanlin Yu, Bernardo Williams et al.

We study a class of optimization problems in the Wasserstein space (the space of probability measures) where the objective function is nonconvex along generalized geodesics. Specifically, the objective exhibits some difference-of-convex structure along these geodesics. The setting also encompasses sampling problems where the logarithm of the target distribution is difference-of-convex. We derive multiple convergence insights for a novel semi Forward-Backward Euler scheme under several nonconvex (and possibly nonsmooth) regimes. Notably, the semi Forward-Backward Euler is just a slight modification of the Forward-Backward Euler whose convergence is -- to our knowledge -- still unknown in our very general non-geodesically-convex setting.