Human Mesh Recovery (HMR) from a single RGB image is a highly ambiguous problem, as an infinite set of 3D interpretations can explain the 2D observation equally well. Nevertheless, most HMR methods overlook this issue and make a single prediction without accounting for this ambiguity. A few approaches generate a distribution of human meshes, enabling the sampling of multiple predictions; however, none of them is competitive with the latest single-output model when making a single prediction. This work proposes a new approach based on masked generative modeling. By tokenizing the human pose and shape, we formulate the HMR task as generating a sequence of discrete tokens conditioned on an input image. We introduce MEGA, a MaskEd Generative Autoencoder trained to recover human meshes from images and partial human mesh token sequences. Given an image, our flexible generation scheme allows us to predict a single human mesh in deterministic mode or to generate multiple human meshes in stochastic mode. Experiments on in-the-wild benchmarks show that MEGA achieves state-of-the-art performance in deterministic and stochastic modes, outperforming single-output and multi-output approaches. See the project page at https://g-fiche.github.io/research-pages/mega/.

This article introduces AnCoGen, a novel method that leverages a masked autoencoder to unify the analysis, control, and generation of speech signals within a single model. AnCoGen can analyze speech by estimating key attributes, such as speaker identity, pitch, content, loudness, signal-to-noise ratio, and clarity index. In addition, it can generate speech from these attributes and allow precise control of the synthesized speech by modifying them. Extensive experiments demonstrated the effectiveness of AnCoGen across speech analysis-resynthesis, pitch estimation, pitch modification, and speech enhancement. Code and audio examples are available online1.

Supervised models for speech enhancement are trained using artificially generated mixtures of clean speech and noise signals. However, the synthetic training conditions may not accurately reflect real-world conditions encountered during testing. This discrepancy can result in poor performance when the test domain significantly differs from the synthetic training domain. To tackle this issue, the UDASE task of the 7th CHiME challenge aimed to leverage real-world noisy speech recordings from the test domain for unsupervised domain adaptation of speech enhancement models. Specifically, this test domain corresponds to the CHiME-5 dataset, characterized by real multi-speaker and conversational speech recordings made in noisy and reverberant domestic environments, for which ground-truth clean speech signals are not available. In this paper, we present the objective and subjective evaluations of the systems that were submitted to the CHiME-7 UDASE task, and we provide an analysis of the results. This analysis reveals a limited correlation between subjective ratings and several supervised nonintrusive performance metrics recently proposed for speech enhancement. Conversely, the results suggest that more traditional intrusive objective metrics can be used for in-domain performance evaluation using the reverberant LibriCHiME-5 dataset developed for the challenge. The subjective evaluation indicates that all systems successfully reduced the background noise, but always at the expense of increased distortion. Out of the four speech enhancement methods evaluated subjectively, only one demonstrated an improvement in overall quality compared to the unprocessed noisy speech, highlighting the difficulty of the task. The tools and audio material created for the CHiME-7 UDASE task are shared with the community.

An important challenge in emotion recognition is to develop methods that can leverage unlabeled training data. In this paper, we propose the VQ-MAE-AV model, a self-supervised multimodal model that leverages masked autoencoders to learn representations of audiovisual speech without labels. The model includes vector quantized variational autoencoders that compress raw audio and visual speech data into discrete tokens. The audiovisual speech tokens are used to train a multimodal masked autoencoder that consists of an encoder–decoder architecture with attention mechanisms. The model is designed to extract both local (i.e., at the frame level) and global (i.e., at the sequence level) representations of audiovisual speech. During self-supervised pre-training, the VQ-MAE-AV model is trained on a large-scale unlabeled dataset of audiovisual speech, for the task of reconstructing randomly masked audiovisual speech tokens and with a contrastive learning strategy. During this pre-training, the encoder learns to extract a representation of audiovisual speech that can be subsequently leveraged for emotion recognition. During the supervised fine-tuning stage, a small classification model is trained on top of the VQ-MAE-AV encoder for an emotion recognition task. The proposed approach achieves state-of-the-art emotion recognition results across several datasets in both controlled and in-the-wild conditions. • We present a self-supervised model for audiovisual speech emotion recognition. • The model operates on discrete audiovisual speech tokens. • A multimodal masked autoencoder with attention fuses the audio and visual modalities. • The model achieves state-of-the-art audiovisual speech emotion recognition results. • Ablation studies reveal the importance of each model component.

One of the main challenges in speech emotion recognition is the lack of large labelled datasets. The progress in speech synthesis allows us to generate reliable and realistic expressive speech. In this work, we propose using a state-of-the-art end-to-end speech emotion conversion model to generate new synthetic data for training speech emotion recognition models. We first evaluate the quality of the converted speech on new unseen datasets, which proves to be on par with the training data. Then, we study the effect of using the synthesized speech as data augmentation. We show that this approach improves the overall performance of emotion recognition models on two different datasets, IEMOCAP and RAVDESS, both in the cases of speaker dependent and independent emotion recognition using a fine-tuned wav2vec 2.0.

High-dimensional data such as natural images or speech signals exhibit some form of regularity, preventing their dimensions from varying independently. This suggests that there exists a lower dimensional latent representation from which the high-dimensional observed data were generated. Uncovering the hidden explanatory features of complex data is the goal of representation learning, and deep latent variable generative models have emerged as promising unsupervised approaches. In particular, the variational autoencoder (VAE) which is equipped with both a generative and an inference model allows for the analysis, transformation, and generation of various types of data. Over the past few years, the VAE has been extended to deal with data that are either multimodal or dynamical (i.e., sequential). In this paper, we present a multimodal and dynamical VAE (MDVAE) applied to unsupervised audiovisual speech representation learning. The latent space is structured to dissociate the latent dynamical factors that are shared between the modalities from those that are specific to each modality. A static latent variable is also introduced to encode the information that is constant over time within an audiovisual speech sequence. The model is trained in an unsupervised manner on an audiovisual emotional speech dataset, in two stages. In the first stage, a vector quantized VAE (VQ-VAE) is learned independently for each modality, without temporal modeling. The second stage consists in learning the MDVAE model on the intermediate representation of the VQ-VAEs before quantization. The disentanglement between static versus dynamical and modality-specific versus modality-common information occurs during this second training stage. Extensive experiments are conducted to investigate how audiovisual speech latent factors are encoded in the latent space of MDVAE. These experiments include manipulating audiovisual speech, audiovisual facial image denoising, and audiovisual speech emotion recognition. The results show that MDVAE effectively combines the audio and visual information in its latent space. They also show that the learned static representation of audiovisual speech can be used for emotion recognition with few labeled data, and with better accuracy compared with unimodal baselines and a state-of-the-art supervised model based on an audiovisual transformer architecture.

The dynamical variational autoencoders (DVAEs) are a family of latent-variable deep generative models that extends the VAE to model a sequence of observed data and a corresponding sequence of latent vectors. In almost all the DVAEs of the literature, the temporal dependencies within each sequence and across the two sequences are modeled with recurrent neural networks. In this paper, we propose to model speech signals with the Hierarchical Transformer DVAE (HiT-DVAE), which is a DVAE with two levels of latent variable (sequence-wise and frame-wise) and in which the temporal dependencies are implemented with the Transformer architecture. We show that HiT-DVAE outperforms several other DVAEs for speech spectrogram modeling, while enabling a simpler training procedure, revealing its high potential for downstream low-level speech processing tasks such as speech enhancement.

This work builds on a previous work on unsupervised speech enhancement using a dynamical variational autoencoder (DVAE) as the clean speech model and non-negative matrix factorization (NMF) as the noise model. We propose to replace the NMF noise model with a deep dynamical generative model (DDGM) depending either on the DVAE latent variables, or on the noisy observations, or on both. This DDGM can be trained in three configurations: noise-agnostic, noise-dependent and noise adaptation after noise-dependent training. Experimental results show that the proposed method achieves competitive performance compared to state-of-the-art unsupervised speech enhancement methods, while the noise-dependent training configuration yields a much more time-efficient inference process.

Pose and motion priors are crucial for recovering realistic and accurate human motion from noisy observations. Substantial progress has been made on pose and shape estimation from images, and recent works showed impressive results using priors to refine frame-wise predictions. However, a lot of motion priors only model transitions between consecutive poses and are used in time-consuming optimization procedures, which is problematic for many applications requiring real-time motion capture. We introduce Motion-DVAE, a motion prior to capture the short-term dependencies of human motion. As part of the dynamical variational autoencoder (DVAE) models family, Motion-DVAE combines the generative capability of VAE models and the temporal modeling of recurrent architectures. Together with Motion-DVAE, we introduce an unsupervised learned denoising method unifying regression- and optimization-based approaches in a single framework for real-time 3D human pose estimation. Experiments show that the proposed approach reaches competitive performance with state-of-the-art methods while being much faster.

Technologies play an increasingly important role in sports and become a real competitive advantage for the athletes who benefit from it. Among them, the use of motion capture is developing in various sports to optimize sporting gestures. Unfortunately, traditional motion capture systems are expensive and constraining. Recently developed computer vision-based approaches also struggle in certain sports, like swimming, due to the aquatic environment. One of the reasons for the gap in performance is the lack of labeled datasets with swimming videos. In an attempt to address this issue, we introduce SwimXYZ, a synthetic dataset of swimming motions and videos. SwimXYZ contains 3.4 million frames annotated with ground truth 2D and 3D joints, as well as 240 sequences of swimming motions in the SMPL parameters format. In addition to making this dataset publicly available, we present use cases for SwimXYZ in swimming stroke clustering and 2D pose estimation.

National audience

Dynamical variational autoencoders (DVAEs) are a class of deep generative models with latent variables, dedicated to model time series of high-dimensional data. DVAEs can be considered as extensions of the variational autoencoder (VAE) that include temporal dependencies between successive observed and/or latent vectors. Previous work has shown the interest of using DVAEs over the VAE for speech spectrograms modeling. Independently, the VAE has been successfully applied to speech enhancement in noise, in an unsupervised noise-agnostic set-up that requires neither noise samples nor noisy speech samples at training time, but only requires clean speech signals. In this paper, we extend these works to DVAE-based single-channel unsupervised speech enhancement, hence exploiting both speech signals unsupervised representation learning and dynamics modeling. We propose an unsupervised speech enhancement algorithm that combines a DVAE speech prior pre-trained on clean speech signals with a noise model based on nonnegative matrix factorization, and we derive a variational expectation-maximization (VEM) algorithm to perform speech enhancement. The algorithm is presented with the most general DVAE formulation and is then applied with three specific DVAE models to illustrate the versatility of the framework. Experimental results show that the proposed DVAE-based approach outperforms its VAE-based counterpart, as well as several supervised and unsupervised noise-dependent baselines, especially when the noise type is unseen during training.

Accepted to Interspeech 2021. arXiv admin note: text overlap with arXiv:2008.12595

This paper presents a generative approach to speech enhancement based on a recurrent variational autoencoder (RVAE). The deep generative speech model is trained using clean speech signals only, and it is combined with a nonnegative matrix factorization noise model for speech enhancement. We propose a variational expectation-maximization algorithm where the encoder of the RVAE is finetuned at test time, to approximate the distribution of the latent variables given the noisy speech observations. Compared with previous approaches based on feed-forward fully-connected architectures, the proposed recurrent deep generative speech model induces a posterior temporal dynamic over the latent variables, which is shown to improve the speech enhancement results.

Variational auto-encoders (VAEs) are deep generative latent variable models that can be used for learning the distribution of complex data. VAEs have been successfully used to learn a probabilistic prior over speech signals, which is then used to perform speech enhancement. One advantage of this generative approach is that it does not require pairs of clean and noisy speech signals at training. In this article, we propose audio-visual variants of VAEs for single-channel and speaker-independent speech enhancement. We develop a conditional VAE (CVAE) where the audio speech generative process is conditioned on visual information of the lip region. At test time, the audio-visual speech generative model is combined with a noise model based on nonnegative matrix factorization, and speech enhancement relies on a Monte Carlo expectation-maximization algorithm. Experiments are conducted with the recently published NTCD-TIMIT dataset as well as the GRID corpus. The results confirm that the proposed audio-visual CVAE effectively fuses audio and visual information, and it improves the speech enhancement performance compared with the audio-only VAE model, especially when the speech signal is highly corrupted by noise. We also show that the proposed unsupervised audio-visual speech enhancement approach outperforms a state-of-the-art supervised deep learning method.

In this paper we address speaker-independent multichannel speech enhancement in unknown noisy environments. Our work is based on a well-established multichannel local Gaussian modeling framework. We propose to use a neural network for modeling the speech spectro-temporal content. The parameters of this supervised model are learned using the framework of variational autoencoders. The noisy recording environment is supposed to be unknown, so the noise spectro-temporal modeling remains unsupervised and is based on non-negative matrix factorization (NMF). We develop a Monte Carlo expectation-maximization algorithm and we experimentally show that the proposed approach outperforms its NMF-based counterpart, where speech is modeled using supervised NMF.

This paper focuses on single-channel semi-supervised speech enhancement. We learn a speaker-independent deep generative speech model using the framework of variational autoencoders. The noise model remains unsupervised because we do not assume prior knowledge of the noisy recording environment. In this context, our contribution is to propose a noise model based on alpha-stable distributions, instead of the more conventional Gaussian non-negative matrix factorization approach found in previous studies. We develop a Monte Carlo expectation-maximization algorithm for estimating the model parameters at test time. Experimental results show the superiority of the proposed approach both in terms of perceptual quality and intelligibility of the enhanced speech signal.

We propose a method using a long short-term memory (LSTM) network to estimate the noise power spectral density (PSD) of single-channel audio signals represented in the short-time Fourier transform (STFT) domain. An LSTM network common to all frequency bands is trained, which processes each frequency band individually by mapping the noisy STFT magnitude sequence to its corresponding noise PSD sequence. Unlike deep-learning-based speech-enhancement methods, which learn the full-band spectral structure of speech segments, the proposed method exploits the sub-band STFT magnitude evolution of noise with long time dependence, in the spirit of the unsupervised noise estimators described in the literature. Speaker- and speech-independent experiments with different types of noise show that the proposed method outperforms the unsupervised estimators, and it generalizes well to noise types that are not present in the training set.

In this study, we propose a novel probabilistic model for separating clean speech signals from noisy mixtures by decomposing the mixture spectra into a structured speech part and a more flexible residual part. The main novelty in our model is that it uses a family of heavy-tailed distributions, so called the α-stable distributions, for modeling the residual signal. We develop an expectation-maximization algorithm for parameter estimation and a Monte Carlo scheme for posterior estimation of the clean speech. Our experiments show that the proposed method outperforms relevant factorization-based algorithms by a significant margin.

This paper presents a Bayesian framework for under-determined audio source separation in multichannel reverberant mixtures. We model the source signals as Student's t latent random variables in a time-frequency domain. The specific structure of musical signals in this domain is exploited by means of a nonnegative matrix factorization model. Conversely, we design the mixing model in the time domain. In addition to leading to an exact representation of the convolutive mixing process, this approach allows us to develop simple probabilistic priors for the mixing filters. Indeed, as those filters correspond to room responses they exhibit a simple characteristic structure in the time domain that can be used to guide their estimation. We also rely on the Student's t distribution for modeling the impulse response of the mixing filters. From this model, we develop a variational inference algorithm in order to perform source separation. The experimental evaluation demonstrates the potential of this approach for separating multichannel reverberant mixtures.

A great number of methods for multichannel audio source separation are based on probabilistic approaches in which the sources are modeled as latent random variables in a Time-Frequency (TF) domain. For reverberant mixtures, it is common to approximate the time-domain convolutive mixing process as being instantaneous in the short-term Fourier transform domain, under a short mixing filters assumption. The TF latent sources are then inferred from the TF mixture observations. In this paper we propose to infer the TF latent sources from the time-domain observations. This approach allows us to exactly model the convolutive mixing process. The inference procedure relies on a variational expectation-maximization algorithm. In significant reverberation conditions, our approach leads to a signal-to-distortion ratio improvement of 5.5 dB compared with the usual TF approximation of the convolutive mixing process.

In this paper, we focus on modeling multichannel audio signals in the short-time Fourier transform domain for the purpose of source separation. We propose a probabilistic model based on a class of heavy-tailed distributions, in which the observed mixtures and the latent sources are jointly modeled by using a certain class of multivariate alpha-stable distributions. As opposed to the conventional Gaussian models, where the observations are constrained to lie just within a few standard deviations from the mean, the proposed heavy-tailed model allows us to account for spurious data or important uncertainties in the model. We develop a Monte Carlo Expectation-Maximization algorithm for inferring the sources from the proposed model. We show that our approach leads to significant performance improvements in audio source separation under corrupted mixtures and in spatial audio object coding.

Incorporating prior knowledge about the sources and/or the mixture is a way to improve under-determined audio source separation performance. A great number of informed source separation techniques concentrate on taking priors on the sources into account, but fewer works have focused on constraining the mixing model. In this paper, we address the problem of underdetermined multichannel audio source separation in reverberant conditions. We target a semi-informed scenario where some room parameters are known. Two probabilistic priors on the frequency response of the mixing filters are proposed. Early reverberation is characterized by an autoregressive model while according to statistical room acoustics results, late reverberation is represented by an autoregressive moving average model. Both reverberation models are defined in the frequency domain. They aim to transcribe the temporal characteristics of the mixing filters into frequency-domain correlations. Our approach leads to a maximum a posteriori estimation of the mixing filters which is achieved thanks to the expectation-maximization algorithm. We experimentally show the superiority of this approach compared with a maximum likelihood estimation of the mixing filters.

In this paper, we propose a new method for singing voice detection based on a Bidirectional Long Short-Term Memory (BLSTM) Recurrent Neural Network (RNN). This classifier is able to take a past and future temporal context into account to decide on the presence/absence of singing voice, thus using the inherent sequential aspect of a short-term feature extraction in a piece of music. The BLSTM-RNN contains several hidden layers, so it is able to extract a simple representation fitted to our task from low-level features. The results we obtain significantly outperform state-of-the-art methods on a common database.