Latent diffusion models have enabled continuous-state diffusion models to handle a variety of datasets, including categorical data. However, most methods rely on fixed pretrained embeddings, limiting the benefits of joint training with the diffusion model. While jointly learning the embedding (via reconstruction loss) and the latent diffusion model (via score matching loss) could enhance performance, our analysis shows that end-to-end training risks embedding collapse, degrading generation quality. To address this issue, we introduce CATDM, a continuous diffusion framework within the embedding space that stabilizes training. We propose a novel objective combining the joint embedding-diffusion variational lower bound with a Consistency-Matching (CM) regularizer, alongside a shifted cosine noise schedule and random dropping strategy. The CM regularizer ensures the recovery of the true data distribution. Experiments on benchmarks show that CATDM mitigates embedding collapse, yielding superior results on FFHQ, LSUN Churches, and LSUN Bedrooms. In particular, CATDM achieves an FID of 6.81 on ImageNet 256 256 with 50 steps. It outperforms non-autoregressive models in machine translation and is on a par with previous methods in text generation. These probabilistic models learn the inverse of a Markov chain that gradually converts data into pure Gaussian noise, using noise-conditioned score functions (i.e., gradients of log density), which are defined only for continuous data. The core concept is to progressively recover the original data distribution using a learned transition kernel. They offer stable and relatively efficient training procedures that contribute to their success. Recent advances, such as consistency models (Song et al., 2023; Kim et al., 2023; Luo et al., 2023), have further enhanced diffusion models by reducing the number of sampling steps, making them more practical for real-world applications.

Recent literature has effectively leveraged diffusion models trained on continuous variables as priors for solving inverse problems. Notably, discrete diffusion models with discrete latent codes have shown strong performance, particularly in modalities suited for discrete compressed representations, such as image and motion generation. However, their discrete and non-differentiable nature has limited their application to inverse problems formulated in continuous spaces. This paper presents a novel method for addressing linear inverse problems by leveraging image-generation models based on discrete diffusion as priors. We overcome these limitations by approximating the true posterior distribution with a variational distribution constructed from categorical distributions and continuous relaxation techniques. Furthermore, we employ a star-shaped noise process to mitigate the drawbacks of traditional discrete diffusion models with absorbing states, demonstrating that our method performs comparably to continuous diffusion techniques. To the best of our knowledge, this is the first approach to use discrete diffusion model-based priors for solving image inverse problems. These models operate by iteratively corrupting data then learning to reverse this corruption process, ultimately generating high-quality samples from noise. In parallel with continuous diffusion models, discrete diffusion models have emerged as a compelling alternative. Building on these advancements, researchers have made significant progress in expanding the application of diffusion models.

Selective attention helps us focus on task-relevant aspects in the constant flood of our sensory input. This constraint in our perception allows us to robustly generalize under distractions and to new compositions of perceivable concepts. Transformers employ a similar notion of attention in their architecture, but representation learning models with transformer backbones like CLIP and DINO often fail to demonstrate robustness and compositionality. We highlight a missing architectural prior: unlike human perception, transformer encodings do not separately attend over individual concepts. In response, we propose Sparo, a read-out mechanism that partitions encodings into separately-attended slots, each produced by a single attention head. Using Sparo with CLIP imparts an inductive bias that the vision and text modalities are different views of a shared compositional world with the same corresponding concepts. Using Sparo, we demonstrate improvements on downstream recognition, robustness, retrieval, and compositionality benchmarks with CLIP (up to +14% for ImageNet, +4% for SugarCrepe), and on nearest neighbors and linear probe for ImageNet with DINO (+3% each). We also showcase a powerful ability to intervene and select individual Sparo concepts to further improve downstream task performance (up from +4% to +9% for SugarCrepe) and use this ability to study the robustness of Sparo's representation structure. Finally, we provide insights through ablation experiments and visualization of learned concepts.

Self-supervised learning (SSL) has achieved remarkable success across various speech-processing tasks. To enhance its efficiency, previous works often leverage the use of compression techniques. A notable recent attempt is DPHuBERT, which applies joint knowledge distillation (KD) and structured pruning to learn a significantly smaller SSL model. In this paper, we contribute to this research domain by introducing SKILL, a novel method that conducts distillation across groups of layers instead of distilling individual arbitrarily selected layers within the teacher network. The identification of the layers to distill is achieved through a hierarchical clustering procedure applied to layer similarity measures. Extensive experiments demonstrate that our distilled version of WavLM Base+ not only outperforms DPHuBERT but also achieves state-of-the-art results in the 30M parameters model class across several SUPERB tasks.

State-of-the-art non-autoregressive text-to-speech (TTS) models based on FastSpeech 2 can efficiently synthesise high-fidelity and natural speech. For expressive speech datasets however, we observe characteristic audio distortions. We demonstrate that such artefacts are introduced to the vocoder reconstruction by over-smooth mel-spectrogram predictions, which are induced by the choice of mean-squared-error (MSE) loss for training the mel-spectrogram decoder. With MSE loss FastSpeech 2 is limited to learn conditional averages of the training distribution, which might not lie close to a natural sample if the distribution still appears multimodal after all conditioning signals. To alleviate this problem, we introduce TVC-GMM, a mixture model of Trivariate-Chain Gaussian distributions, to model the residual multimodality. TVC-GMM reduces spectrogram smoothness and improves perceptual audio quality in particular for expressive datasets as shown by both objective and subjective evaluation.

Deep equilibrium models (DEQs) have proven to be very powerful for learning data representations. The idea is to replace traditional (explicit) feedforward neural networks with an implicit fixed-point equation, which allows to decouple the forward and backward passes. In particular, training DEQ layers becomes very memory-efficient via the implicit function theorem. However, backpropagation through DEQ layers still requires solving an expensive Jacobian-based equation. In this paper, we introduce a simple but effective strategy to avoid this computational burden. Our method relies on the Jacobian approximation of Broyden's method after the forward pass to compute the gradients during the backward pass. Experiments show that simply re-using this approximation can significantly speed up the training while not causing any performance degradation.

Parallel text-to-speech (TTS) models have recently enabled fast and highly-natural speech synthesis. However, they typically require external alignment models, which are not necessarily optimized for the decoder as they are not jointly trained. In this paper, we propose a differentiable duration method for learning monotonic alignments between input and output sequences. Our method is based on a soft-duration mechanism that optimizes a stochastic process in expectation. Using this differentiable duration method, we introduce AutoTTS, a direct text-to-waveform speech synthesis model. AutoTTS enables high-fidelity speech synthesis through a combination of adversarial training and matching the total ground-truth duration. Experimental results show that our model obtains competitive results while enjoying a much simpler training pipeline. Audio samples are available online.

Voice conversion has gained increasing popularity in many applications of speech synthesis. The idea is to change the voice identity from one speaker into another while keeping the linguistic content unchanged. Many voice conversion approaches rely on the use of a vocoder to reconstruct the speech from acoustic features, and as a consequence, the speech quality heavily depends on such a vocoder. In this paper, we propose NVC-Net, an end-to-end adversarial network, which performs voice conversion directly on the raw audio waveform of arbitrary length. By disentangling the speaker identity from the speech content, NVC-Net is able to perform non-parallel traditional many-to-many voice conversion as well as zero-shot voice conversion from a short utterance of an unseen target speaker. Importantly, NVC-Net is non-autoregressive and fully convolutional, achieving fast inference. Our model is capable of producing samples at a rate of more than 3600 kHz on an NVIDIA V100 GPU, being orders of magnitude faster than state-of-the-art methods under the same hardware configurations. Objective and subjective evaluations on non-parallel many-to-many voice conversion tasks show that NVC-Net obtains competitive results with significantly fewer parameters.