Mostafa Sadeghi

Ali Golmakani, Mostafa Sadeghi, Xavier Alameda-Pineda et al.

We address speech enhancement based on variational autoencoders, which involves learning a speech prior distribution in the time-frequency (TF) domain. A zero-mean complex-valued Gaussian distribution is usually assumed for the generative model, where the speech information is encoded in the variance as a function of a latent variable. In contrast to this commonly used approach, we propose a weighted variance generative model, where the contribution of each spectrogram time-frame in parameter learning is weighted. We impose a Gamma prior distribution on the weights, which would effectively lead to a Student's t-distribution instead of Gaussian for speech generative modeling. We develop efficient training and speech enhancement algorithms based on the proposed generative model. Our experimental results on spectrogram auto-encoding and speech enhancement demonstrate the effectiveness and robustness of the proposed approach compared to the standard unweighted variance model.

Ali Golmakani, Mostafa Sadeghi, Romain Serizel

Deep latent variable generative models based on variational autoencoder (VAE) have shown promising performance for audiovisual speech enhancement (AVSE). The underlying idea is to learn a VAEbased audiovisual prior distribution for clean speech data, and then combine it with a statistical noise model to recover a speech signal from a noisy audio recording and video (lip images) of the target speaker. Existing generative models developed for AVSE do not take into account the sequential nature of speech data, which prevents them from fully incorporating the power of visual data. In this paper, we present an audiovisual deep Kalman filter (AV-DKF) generative model which assumes a first-order Markov chain model for the latent variables and effectively fuses audiovisual data. Moreover, we develop an efficient inference methodology to estimate speech signals at test time. We conduct a set of experiments to compare different variants of generative models for speech enhancement. The results demonstrate the superiority of the AV-DKF model compared with both its audio-only version and the non-sequential audio-only and audiovisual VAE-based models.

Mostafa Sadeghi, Romain Serizel

Unsupervised speech enhancement based on variational autoencoders has shown promising performance compared with the commonly used supervised methods. This approach involves the use of a pre-trained deep speech prior along with a parametric noise model, where the noise parameters are learned from the noisy speech signal with an expectationmaximization (EM)-based method. The E-step involves an intractable latent posterior distribution. Existing algorithms to solve this step are either based on computationally heavy Monte Carlo Markov Chain sampling methods and variational inference, or inefficient optimization-based methods. In this paper, we propose a new approach based on Langevin dynamics that generates multiple sequences of samples and comes with a total variation-based regularization to incorporate temporal correlations of latent vectors. Our experiments demonstrate that the developed framework makes an effective compromise between computational efficiency and enhancement quality, and outperforms existing methods.

Berné Nortier, Mostafa Sadeghi, Romain Serizel

Recently, conditional score-based diffusion models have gained significant attention in the field of supervised speech enhancement, yielding state-of-the-art performance. However, these methods may face challenges when generalising to unseen conditions. To address this issue, we introduce an alternative approach that operates in an unsupervised manner, leveraging the generative power of diffusion models. Specifically, in a training phase, a clean speech prior distribution is learnt in the short-time Fourier transform (STFT) domain using score-based diffusion models, allowing it to unconditionally generate clean speech from Gaussian noise. Then, we develop a posterior sampling methodology for speech enhancement by combining the learnt clean speech prior with a noise model for speech signal inference. The noise parameters are simultaneously learnt along with clean speech estimation through an iterative expectationmaximisation (EM) approach. To the best of our knowledge, this is the first work exploring diffusion-based generative models for unsupervised speech enhancement, demonstrating promising results compared to a recent variational auto-encoder (VAE)-based unsupervised approach and a state-of-the-art diffusion-based supervised method. It thus opens a new direction for future research in unsupervised speech enhancement.

Can Cui, Imran Ahamad Sheikh, Mostafa Sadeghi et al.

We present an end-to-end multichannel speaker-attributed automatic speech recognition (MC-SA-ASR) system that combines a Conformer-based encoder with multi-frame crosschannel attention and a speaker-attributed Transformer-based decoder. To the best of our knowledge, this is the first model that efficiently integrates ASR and speaker identification modules in a multichannel setting. On simulated mixtures of LibriSpeech data, our system reduces the word error rate (WER) by up to 12% and 16% relative compared to previously proposed single-channel and multichannel approaches, respectively. Furthermore, we investigate the impact of different input features, including multichannel magnitude and phase information, on the ASR performance. Finally, our experiments on the AMI corpus confirm the effectiveness of our system for real-world multichannel meeting transcription.

Can Cui, Imran Ahamad Sheikh, Mostafa Sadeghi et al.

Joint punctuated and normalized automatic speech recognition (ASR) aims at outputing transcripts with and without punctuation and casing. This task remains challenging due to the lack of paired speech and punctuated text data in most ASR corpora. We propose two approaches to train an end-to-end joint punctuated and normalized ASR system using limited punctuated data. The first approach uses a language model to convert normalized training transcripts into punctuated transcripts. This achieves a better performance on out-of-domain test data, with up to 17% relative Punctuation-Case-aware Word Error Rate (PC-WER) reduction. The second approach uses a single decoder conditioned on the type of output. This yields a 42% relative PC-WER reduction compared to Whisper-base and a 4% relative (normalized) WER reduction compared to the normalized output of a punctuated-only model. Additionally, our proposed model demonstrates the feasibility of a joint ASR system using as little as 5% punctuated training data with a moderate (2.42% absolute) PC-WER increase.

Mostafa Sadeghi, Romain Serizel

In this paper, we address the unsupervised speech enhancement problem based on recurrent variational autoencoder (RVAE). This approach offers promising generalization performance over the supervised counterpart. Nevertheless, the involved iterative variational expectation-maximization (VEM) process at test time, which relies on a variational inference method, results in high computational complexity. To tackle this issue, we present efficient sampling techniques based on Langevin dynamics and Metropolis-Hasting algorithms, adapted to the EM-based speech enhancement with RVAE. By directly sampling from the intractable posterior distribution within the EM process, we circumvent the intricacies of variational inference. We conduct a series of experiments, comparing the proposed methods with VEM and a state-of-the-art supervised speech enhancement approach based on diffusion models. The results reveal that our sampling-based algorithms significantly outperform VEM, not only in terms of computational efficiency but also in overall performance. Furthermore, when compared to the supervised baseline, our methods showcase robust generalization performance in mismatched test conditions.

Guilhem Fauré, Mostafa Sadeghi, Sam Bigeard et al.

Latent diffusion approaches to sign language production (SLP) rely on an initial stage that learns an encoding of sign pose sequences, enabling generative modeling in the resulting latent space. The autoencoder used in this stage is typically evaluated in terms of reconstruction quality using geometric metrics common in SLP. While informative, these metrics do not fully capture latent space properties that may influence the training and performance of the downstream generative model. In this work, we investigate how architectural and training objective design choices in a variational autoencoder (VAE) for sign pose encoding affect latent space structure, and how these differences translate into the performance of a latent diffusion model for text-to-sign generation. Our experiments on Phoenix14T dataset show that variations in generative performance, measured through back-translation BLEU scores, can sometimes be better explained by differences in latent space properties than by VAE reconstruction accuracy alone.

Can Cui, Imran Ahamad Sheikh, Mostafa Sadeghi et al.

Past studies on end-to-end meeting transcription have focused on model architecture and have mostly been evaluated on simulated meeting data. We present a novel study aiming to optimize the use of a Speaker-Attributed ASR (SA-ASR) system in real-life scenarios, such as the AMI meeting corpus, for improved speaker assignment of speech segments. First, we propose a pipeline tailored to real-life applications involving Voice Activity Detection (VAD), Speaker Diarization (SD), and SA-ASR. Second, we advocate using VAD output segments to fine-tune the SA-ASR model, considering that it is also applied to VAD segments during test, and show that this results in a relative reduction of Speaker Error Rate (SER) up to 28%. Finally, we explore strategies to enhance the extraction of the speaker embedding templates used as inputs by the SA-ASR system. We show that extracting them from SD output rather than annotated speaker segments results in a relative SER reduction up to 20%.

Can Cui, Imran Ahamad Sheikh, Mostafa Sadeghi et al.

Distant-microphone meeting transcription is a challenging task. State-of-the-art end-to-end speaker-attributed automatic speech recognition (SA-ASR) architectures lack a multichannel noise and reverberation reduction front-end, which limits their performance. In this paper, we introduce a joint beamforming and SA-ASR approach for real meeting transcription. We first describe a data alignment and augmentation method to pretrain a neural beamformer on real meeting data. We then compare fixed, hybrid, and fully neural beamformers as front-ends to the SA-ASR model. Finally, we jointly optimize the fully neural beamformer and the SA-ASR model. Experiments on the real AMI corpus show that, while state-of-the-art multi-frame cross-channel attention based channel fusion fails to improve ASR performance, fine-tuning SA-ASR on the fixed beamformer's output and jointly fine-tuning SA-ASR with the neural beamformer reduce the word error rate by 8% and 9% relative, respectively.

Guilhem Fauré, Mostafa Sadeghi, Sam Bigeard et al.

One of the main challenges in neural sign language production (SLP) lies in the high intra-class variability of signs, arising from signer morphology and stylistic variety in the training data. To improve robustness to such variations, we propose two enhancements to the standard Progressive Transformers (PT) architecture (Saunders et al., 2020). First, we encode poses using bone rotations in quaternion space and train with a geodesic loss to improve the accuracy and clarity of angular joint movements. Second, we introduce a contrastive loss to structure decoder embeddings by semantic similarity, using either gloss overlap or SBERT-based sentence similarity, aiming to filter out anatomical and stylistic features that do not convey relevant semantic information. On the Phoenix14T dataset, the contrastive loss alone yields a 16% improvement in Probability of Correct Keypoint over the PT baseline. When combined with quaternion-based pose encoding, the model achieves a 6% reduction in Mean Bone Angle Error. These results point to the benefit of incorporating skeletal structure modeling and semantically guided contrastive objectives on sign pose representations into the training of Transformer-based SLP models.

Mostafa Sadeghi, Paul Magron

Structuring the latent space in probabilistic deep generative models, e.g., variational autoencoders (VAEs), is important to yield more expressive models and interpretable representations, and to avoid overfitting. One way to achieve this objective is to impose a sparsity constraint on the latent variables, e.g., via a Laplace prior. However, such approaches usually complicate the training phase, and they sacrifice the reconstruction quality to promote sparsity. In this paper, we propose a simple yet effective methodology to structure the latent space via a sparsity-promoting dictionary model, which assumes that each latent code can be written as a sparse linear combination of a dictionary's columns. In particular, we leverage a computationally efficient and tuning-free method, which relies on a zero-mean Gaussian latent prior with learnable variances. We derive a variational inference scheme to train the model. Experiments on speech generative modeling demonstrate the advantage of the proposed approach over competing techniques, since it promotes sparsity while not deteriorating the output speech quality.

Zhiqi Kang, Mostafa Sadeghi, Radu Horaud et al.

This paper investigates the impact of head movements on audio-visual speech enhancement (AVSE). Although being a common conversational feature, head movements have been ignored by past and recent studies: they challenge today's learning-based methods as they often degrade the performance of models that are trained on clean, frontal, and steady face images. To alleviate this problem, we propose to use robust face frontalization (RFF) in combination with an AVSE method based on a variational auto-encoder (VAE) model. We briefly describe the basic ingredients of the proposed pipeline and we perform experiments with a recently released audio-visual dataset. In the light of these experiments, and based on three standard metrics, namely STOI, PESQ and SI-SDR, we conclude that RFF improves the performance of AVSE by a considerable margin.

Mostafa Sadeghi, Xavier Alameda-Pineda

Recently, audio-visual speech enhancement has been tackled in the unsupervised settings based on variational auto-encoders (VAEs), where during training only clean data is used to train a generative model for speech, which at test time is combined with a noise model, e.g. nonnegative matrix factorization (NMF), whose parameters are learned without supervision. Consequently, the proposed model is agnostic to the noise type. When visual data are clean, audio-visual VAE-based architectures usually outperform the audio-only counterpart. The opposite happens when the visual data are corrupted by clutter, e.g. the speaker not facing the camera. In this paper, we propose to find the optimal combination of these two architectures through time. More precisely, we introduce the use of a latent sequential variable with Markovian dependencies to switch between different VAE architectures through time in an unsupervised manner: leading to switching variational auto-encoder (SwVAE). We propose a variational factorization to approximate the computationally intractable posterior distribution. We also derive the corresponding variational expectation-maximization algorithm to estimate the parameters of the model and enhance the speech signal. Our experiments demonstrate the promising performance of SwVAE.

Mostafa Sadeghi, Xavier Alameda-Pineda

In this paper, we are interested in unsupervised (unknown noise) audio-visual speech enhancement based on variational autoencoders (VAEs), where the probability distribution of clean speech spectra is simulated using an encoder-decoder architecture. The trained generative model (decoder) is then combined with a noise model at test time to estimate the clean speech. In the speech enhancement phase (test time), the initialization of the latent variables, which describe the generative process of clean speech via decoder, is crucial, as the overall inference problem is non-convex. This is usually done by using the output of the trained encoder where the noisy audio and clean visual data are given as input. Current audio-visual VAE models do not provide an effective initialization because the two modalities are tightly coupled (concatenated) in the associated architectures. To overcome this issue, inspired by mixture models, we introduce the mixture of inference networks variational autoencoder (MIN-VAE). Two encoder networks input, respectively, audio and visual data, and the posterior of the latent variables is modeled as a mixture of two Gaussian distributions output from each encoder network. The mixture variable is also latent, and therefore the inference of learning the optimal balance between the audio and visual inference networks is unsupervised as well. By training a shared decoder, the overall network learns to adaptively fuse the two modalities. Moreover, at test time, the visual encoder, which takes (clean) visual data, is used for initialization. A variational inference approach is derived to train the proposed generative model. Thanks to the novel inference procedure and the robust initialization, the proposed MIN-VAE exhibits superior performance on speech enhancement than using the standard audio-only as well as audio-visual counterparts.

Mostafa Sadeghi, Xavier Alameda-Pineda

Recently, an audio-visual speech generative model based on variational autoencoder (VAE) has been proposed, which is combined with a nonnegative matrix factorization (NMF) model for noise variance to perform unsupervised speech enhancement. When visual data is clean, speech enhancement with audio-visual VAE shows a better performance than with audio-only VAE, which is trained on audio-only data. However, audio-visual VAE is not robust against noisy visual data, e.g., when for some video frames, speaker face is not frontal or lips region is occluded. In this paper, we propose a robust unsupervised audio-visual speech enhancement method based on a per-frame VAE mixture model. This mixture model consists of a trained audio-only VAE and a trained audio-visual VAE. The motivation is to skip noisy visual frames by switching to the audio-only VAE model. We present a variational expectation-maximization method to estimate the parameters of the model. Experiments show the promising performance of the proposed method.

Mostafa Sadeghi, Simon Leglaive, Xavier Alameda-PIneda et al.

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 paper, 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.

Saikat Chatterjee, Alireza M. Javid, Mostafa Sadeghi et al.

We design a self size-estimating feed-forward network (SSFN) using a joint optimization approach for estimation of number of layers, number of nodes and learning of weight matrices. The learning algorithm has a low computational complexity, preferably within few minutes using a laptop. In addition the algorithm has a limited need for human intervention to tune parameters. SSFN grows from a small-size network to a large-size network, guaranteeing a monotonically non-increasing cost with addition of nodes and layers. The learning approach uses judicious a combination of `lossless flow property' of some activation functions, convex optimization and instance of random matrix. Consistent performance -- low variation across Monte-Carlo trials -- is found for inference performance (classification accuracy) and estimation of network size.

Saikat Chatterjee, Alireza M. Javid, Mostafa Sadeghi et al.

We develop an algorithm for systematic design of a large artificial neural network using a progression property. We find that some non-linear functions, such as the rectifier linear unit and its derivatives, hold the property. The systematic design addresses the choice of network size and regularization of parameters. The number of nodes and layers in network increases in progression with the objective of consistently reducing an appropriate cost. Each layer is optimized at a time, where appropriate parameters are learned using convex optimization. Regularization parameters for convex optimization do not need a significant manual effort for tuning. We also use random instances for some weight matrices, and that helps to reduce the number of parameters we learn. The developed network is expected to show good generalization power due to appropriate regularization and use of random weights in the layers. This expectation is verified by extensive experiments for classification and regression problems, using standard databases.