Generative speech enhancement has recently shown promising advancements in improving speech quality in noisy environments. Multiple diffusion-based frameworks exist, each employing distinct training objectives and learning techniques. This paper aims to explain the differences between these frameworks by focusing our investigation on score-based generative models and the Schrodinger bridge. We conduct a series of comprehensive experiments to compare their performance and highlight differing training behaviors. Furthermore, we propose a novel perceptual loss function tailored for the Schrodinger bridge framework, demonstrating enhanced performance and improved perceptual quality of the enhanced speech signals. All experimental code and pre-trained models are publicly available to facilitate further research and development in this domain1.

Diffusion models have shown promising results in single-channel speech enhancement, using a task-adapted diffusion process for the conditional generation of clean speech given a noisy mixture. However, at test time, the neural network used for score estimation is called multiple times to solve the iterative reverse process. This results in a slow inference process and causes discretization errors that accumulate over the sampling trajectory. In this paper, we address these limitations through a two-stage training approach. In the first stage, we train the diffusion model the usual way using the generative denoising score matching loss. In the second stage, we compute the enhanced signal by solving the reverse process and compare the resulting estimate to the clean speech target using a predictive loss. We show that using this second training stage enables achieving the same performance as the baseline model using only 5 function evaluations instead of 60 function evaluations. While the performance of usual generative diffusion algorithms drops dramatically when lowering the number of function evaluations to obtain single-step diffusion, we show that our proposed method keeps a steady performance and therefore largely outperforms the diffusion baseline in this setting and also generalizes better than its predictive counterpart1.

Diffusion-based generative models have had a high impact on the computer vision and speech processing communities these past years. Besides data generation tasks, they have also been employed for data restoration tasks like speech enhancement and dereverberation. While discriminative models have traditionally been argued to be more powerful e.g. for speech enhancement, generative diffusion approaches have recently been shown to narrow this performance gap considerably. In this paper, we systematically compare the performance of generative diffusion models and discriminative approaches on different speech restoration tasks. For this, we extend our prior contributions on diffusion-based speech enhancement in the complex time-frequency domain to the task of bandwith extension. We then compare it to a discriminatively trained neural network with the same network architecture on three restoration tasks, namely speech denoising, dereverberation and bandwidth extension. We observe that the generative approach performs globally better than its discriminative counterpart on all tasks, with the strongest benefit for non-additive distortion models, like in dereverberation and bandwidth extension. Code and audio examples can be found online1.

In this paper, we present a causal speech signal improvement system that is designed to handle different types of distortions. The method is based on a generative diffusion model which has been shown to work well in scenarios with missing data and non-linear corruptions. To guarantee causal processing, we modify the network architecture of our previous work and replace global normalization with causal adaptive gain control. We generate diverse training data containing a broad range of distortions. This work was performed in the context of an "ICASSP Signal Processing Grand Challenge" and submitted to the non-real-time track of the "Speech Signal Improvement Challenge 2023", where it was ranked fifth.

Diffusion models have shown a great ability at bridging the performance gap between predictive and generative approaches for speech enhancement. We have shown that they may even outperform their predictive counterparts for non-additive corruption types or when they are evaluated on mismatched conditions. However, diffusion models suffer from a high computational burden, mainly as they require to run a neural network for each reverse diffusion step, whereas predictive approaches only require one pass. As diffusion models are generative approaches they may also produce vocalizing and breathing artifacts in adverse conditions. In comparison, in such difficult scenarios, predictive models typically do not produce such artifacts but tend to distort the target speech instead, thereby degrading the speech quality. In this work, we present a stochastic regeneration approach where an estimate given by a predictive model is provided as a guide for further diffusion. We show that the proposed approach uses the predictive model to remove the vocalizing and breathing artifacts while producing very high quality samples thanks to the diffusion model, even in adverse conditions. We further show that this approach enables to use lighter sampling schemes with fewer diffusion steps without sacrificing quality, thus lifting the computational burden by an order of magnitude. Source code and audio examples are available onlinehttps://uhh.de/inf-sp-storm.

In this work, we build upon our previous publication and use diffusion-based generative models for speech enhancement. We present a detailed overview of the diffusion process that is based on a stochastic differential equation and delve into an extensive theoretical examination of its implications. Opposed to usual conditional generation tasks, we do not start the reverse process from pure Gaussian noise but from a mixture of noisy speech and Gaussian noise. This matches our forward process which moves from clean speech to noisy speech by including a drift term. We show that this procedure enables using only 30 diffusion steps to generate high-quality clean speech estimates. By adapting the network architecture, we are able to significantly improve the speech enhancement performance, indicating that the network, rather than the formalism, was the main limitation of our original approach. In an extensive cross-dataset evaluation, we show that the improved method can compete with recent discriminative models and achieves better generalization when evaluating on a different corpus than used for training. We complement the results with an instrumental evaluation using real-world noisy recordings and a listening experiment, in which our proposed method is rated best. Examining different sampler configurations for solving the reverse process allows us to balance the performance and computational speed of the proposed method. Moreover, we show that the proposed method is also suitable for dereverberation and thus not limited to additive background noise removal. Code and audio examples are available online1https://github.com/sp-uhh/sgmse.

While state-of-the-art audio-only phoneme recognition is already at a high standard, the robustness of existing methods still drops in very noisy environments. To mitigate these limitations, visual information can be incorporated into the recognition system, such that the problem is formulated in a multi-modal setting. To this end, we develop a continuous, audio-visual phoneme classifier that takes raw audio waveforms and video frames as input. Both modalities are processed by individual feature extraction models before a fusion model exploits their correlations. Audio features are extracted with a residual neural network, while video features are obtained with a convolutional neural network. Furthermore, we model temporal dependencies with gated recurrent units. For modality fusion, we compare simple concatenation, attention-based methods, as well as squeeze-and-excitation to learn a joint representation. We train our models on the NTCD-TIMIT dataset, using distinct noise types from the QUT dataset for the test. By pre-training the feature extraction models on the individual modalities first, we achieve best performance for the audio-visual model that is trained end-to-end. In the experiments, we show that by including the video modality, we increase the accuracy of phoneme prediction by 9% in very noisy acoustic environments. The results indicate that in such environments our approach remains more robust compared to existing methods. The code and pre-trained models are available online11https://github.com/sp-uhh/av-phoneme.

Score-based generative models (SGMs) have recently shown impressive results for difficult generative tasks such as the unconditional and conditional generation of natural images and audio signals.In this work, we extend these models to the complex short-time Fourier transform (STFT) domain, proposing a novel training task for speech enhancement using a complex-valued deep neural network.We derive this training task within the formalism of stochastic differential equations (SDEs), thereby enabling the use of predictor-corrector samplers.We provide alternative formulations inspired by previous publications on using generative diffusion models for speech enhancement, avoiding the need for any prior assumptions on the noise distribution and making the training task purely generative which, as we show, results in improved enhancement performance.

Recently, variational autoencoders have been successfully used to learn a probabilistic prior over speech signals, which is then used to perform speech enhancement. However, variational autoencoders are trained on clean speech only, which results in a limited ability of extracting the speech signal from noisy speech compared to supervised approaches. In this paper, we propose to guide the variational autoencoder with a supervised classifier separately trained on noisy speech. The estimated label is a high-level categorical variable describing the speech signal (e.g. speech activity) allowing for a more informed latent distribution compared to the standard variational autoencoder. We evaluate our method with different types of labels on real recordings of different noisy environments. Provided that the label better informs the latent distribution and that the classifier achieves good performance, the proposed approach outperforms the standard variational autoencoder and a conventional neural network- based supervised approach.

The performance of an audio-visual sound source separation system is determined by its ability to separate audio sources given the images of the sources and the audio mixture. The goal of this study is to investigate the ability to learn the mapping between the sounds and the images of instruments in the self-supervisied mix-and-seperate training paradigm used by state-of-the-art audio-visual sound source separation methods. Theoretical and empirical analyses illustrate that the self-supervised mix-and-separate training does not automatically learn the 1-to-1 correspondence between visual and audio signals, leading to low audio-visual object classification accuracy. Based on this analysis, a weakly-supervised method called Object-Prior is proposed and evaluated on two audio-visual datasets. The experimental results show that the Object-Prior method outperforms state-of-the-art baselines in the audio-visual sound source separation task. It is also more robust against asynchronized data, where the frame and the audio do not come from the same video, and recognizes musical instruments based on their sound with higher accuracy. This indicates that learning the 1-to-1 correspondence between visual and audio features of an instrument improves the effectiveness of audio-visual sound source separation.