Timo Gerkmann

Julius Richter, Simon Welker, Jean-Marie Lemercier et al.

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 online, see https://github.com/sp-uhh/sgmse.

Simon Welker, Henry N. Chapman, Timo Gerkmann

In this work, we utilize the high-fidelity generation abilities of diffusion models to solve blind JPEG restoration at high compression levels. We propose an elegant modification of the forward stochastic differential equation of diffusion models to adapt them to this restoration task and name our method DriftRec. Comparing DriftRec against an $L_2$ regression baseline with the same network architecture and state-of-the-art techniques for JPEG restoration, we show that our approach can escape the tendency of other methods to generate blurry images, and recovers the distribution of clean images significantly more faithfully. For this, only a dataset of clean/corrupted image pairs and no knowledge about the corruption operation is required, enabling wider applicability to other restoration tasks. In contrast to other conditional and unconditional diffusion models, we utilize the idea that the distributions of clean and corrupted images are much closer to each other than each is to the usual Gaussian prior of the reverse process in diffusion models. Our approach therefore requires only low levels of added noise and needs comparatively few sampling steps even without further optimizations. We show that DriftRec naturally generalizes to realistic and difficult scenarios such as unaligned double JPEG compression and blind restoration of JPEGs found online, without having encountered such examples during training.

Jean-Marie Lemercier, Julius Richter, Simon Welker et al.

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 online (https://uhh.de/inf-sp-storm).

Kristina Tesch, Timo Gerkmann

The key advantage of using multiple microphones for speech enhancement is that spatial filtering can be used to complement the tempo-spectral processing. In a traditional setting, linear spatial filtering (beamforming) and single-channel post-filtering are commonly performed separately. In contrast, there is a trend towards employing deep neural networks (DNNs) to learn a joint spatial and tempo-spectral non-linear filter, which means that the restriction of a linear processing model and that of a separate processing of spatial and tempo-spectral information can potentially be overcome. However, the internal mechanisms that lead to good performance of such data-driven filters for multi-channel speech enhancement are not well understood. Therefore, in this work, we analyse the properties of a non-linear spatial filter realized by a DNN as well as its interdependency with temporal and spectral processing by carefully controlling the information sources (spatial, spectral, and temporal) available to the network. We confirm the superiority of a non-linear spatial processing model, which outperforms an oracle linear spatial filter in a challenging speaker extraction scenario for a low number of microphones by 0.24 POLQA score. Our analyses reveal that in particular spectral information should be processed jointly with spatial information as this increases the spatial selectivity of the filter. Our systematic evaluation then leads to a simple network architecture, that outperforms state-of-the-art network architectures on a speaker extraction task by 0.22 POLQA score and by 0.32 POLQA score on the CHiME3 data.

Jean-Marie Lemercier, Julius Richter, Simon Welker et al.

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 online at https://uhh.de/inf-sp-sgmsemultitask

Bunlong Lay, Simon Welker, Julius Richter et al.

Recently, score-based generative models have been successfully employed for the task of speech enhancement. A stochastic differential equation is used to model the iterative forward process, where at each step environmental noise and white Gaussian noise are added to the clean speech signal. While in limit the mean of the forward process ends at the noisy mixture, in practice it stops earlier and thus only at an approximation of the noisy mixture. This results in a discrepancy between the terminating distribution of the forward process and the prior used for solving the reverse process at inference. In this paper, we address this discrepancy and propose a forward process based on a Brownian bridge. We show that such a process leads to a reduction of the mismatch compared to previous diffusion processes. More importantly, we show that our approach improves in objective metrics over the baseline process with only half of the iteration steps and having one hyperparameter less to tune.

Kristina Tesch, Timo Gerkmann

In a multi-channel separation task with multiple speakers, we aim to recover all individual speech signals from the mixture. In contrast to single-channel approaches, which rely on the different spectro-temporal characteristics of the speech signals, multi-channel approaches should additionally utilize the different spatial locations of the sources for a more powerful separation especially when the number of sources increases. To enhance the spatial processing in a multi-channel source separation scenario, in this work, we propose a deep neural network (DNN) based spatially selective filter (SSF) that can be spatially steered to extract the speaker of interest by initializing a recurrent neural network layer with the target direction. We compare the proposed SSF with a common end-to-end direct separation (DS) approach trained using utterance-wise permutation invariant training (PIT), which only implicitly learns to perform spatial filtering. We show that the SSF has a clear advantage over a DS approach with the same underlying network architecture when there are more than two speakers in the mixture, which can be attributed to a better use of the spatial information. Furthermore, we find that the SSF generalizes much better to additional noise sources that were not seen during training and to scenarios with speakers positioned at a similar angle.

Jean-Marie Lemercier, Eloi Moliner, Simon Welker et al.

This paper presents an unsupervised method for single-channel blind dereverberation and room impulse response (RIR) estimation, called BUDDy. The algorithm is rooted in Bayesian posterior sampling: it combines a likelihood model enforcing fidelity to the reverberant measurement, and an anechoic speech prior implemented by an unconditional diffusion model. We design a parametric filter representing the RIR, with exponential decay for each frequency subband. Room acoustics estimation and speech dereverberation are jointly carried out, as the filter parameters are iteratively estimated and the speech utterance refined along the reverse diffusion trajectory. In a blind scenario where the RIR is unknown, BUDDy successfully performs speech dereverberation in various acoustic scenarios, significantly outperforming other blind unsupervised baselines. Unlike supervised methods, which often struggle to generalize, BUDDy seamlessly adapts to different acoustic conditions. This paper extends our previous work by offering new experimental results and insights into the algorithm's versatility. We demonstrate the robustness of our proposed method to new acoustic and speaker conditions, as well as its adaptability to high-resolution singing voice dereverberation, using both instrumental metrics and subjective listening evaluation. We study BUDDy's performance for RIR estimation and observe it surpasses a state-of-the-art supervised DNN-based estimator on mismatched acoustic conditions. Finally, we investigate the sensitivity of informed dereverberation methods to RIR estimation errors, thereby motivating the joint acoustic estimation and dereverberation design. Audio examples and code can be found online.

Kristina Tesch, Timo Gerkmann

In a scenario with multiple persons talking simultaneously, the spatial characteristics of the signals are the most distinct feature for extracting the target signal. In this work, we develop a deep joint spatial-spectral non-linear filter that can be steered in an arbitrary target direction. For this we propose a simple and effective conditioning mechanism, which sets the initial state of the filter's recurrent layers based on the target direction. We show that this scheme is more effective than the baseline approach and increases the flexibility of the filter at no performance cost. The resulting spatially selective non-linear filters can also be used for speech separation of an arbitrary number of speakers and enable very accurate multi-speaker localization as we demonstrate in this paper.

Danilo de Oliveira, Tal Peer, Timo Gerkmann

The SepFormer architecture shows very good results in speech separation. Like other learned-encoder models, it uses short frames, as they have been shown to obtain better performance in these cases. This results in a large number of frames at the input, which is problematic; since the SepFormer is transformer-based, its computational complexity drastically increases with longer sequences. In this paper, we employ the SepFormer in a speech enhancement task and show that by replacing the learned-encoder features with a magnitude short-time Fourier transform (STFT) representation, we can use long frames without compromising perceptual enhancement performance. We obtained equivalent quality and intelligibility evaluation scores while reducing the number of operations by a factor of approximately 8 for a 10-second utterance.

Bunlong Lay, Jean-Marie Lemercier, Julius Richter et al.

Diffusion models have shown promising results in 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 (NFEs) 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 counterpart.

Tal Peer, Simon Welker, Timo Gerkmann

Diffusion probabilistic models have been recently used in a variety of tasks, including speech enhancement and synthesis. As a generative approach, diffusion models have been shown to be especially suitable for imputation problems, where missing data is generated based on existing data. Phase retrieval is inherently an imputation problem, where phase information has to be generated based on the given magnitude. In this work we build upon previous work in the speech domain, adapting a speech enhancement diffusion model specifically for STFT phase retrieval. Evaluation using speech quality and intelligibility metrics shows the diffusion approach is well-suited to the phase retrieval task, with performance surpassing both classical and modern methods.

Julius Richter, Simone Frintrop, Timo Gerkmann

This paper introduces an audio-visual speech enhancement system that leverages score-based generative models, also known as diffusion models, conditioned on visual information. In particular, we exploit audio-visual embeddings obtained from a self-super\-vised learning model that has been fine-tuned on lipreading. The layer-wise features of its transformer-based encoder are aggregated, time-aligned, and incorporated into the noise conditional score network. Experimental evaluations show that the proposed audio-visual speech enhancement system yields improved speech quality and reduces generative artifacts such as phonetic confusions with respect to the audio-only equivalent. The latter is supported by the word error rate of a downstream automatic speech recognition model, which decreases noticeably, especially at low input signal-to-noise ratios.

Huajian Fang, Tal Peer, Stefan Wermter et al.

Speech enhancement in the time-frequency domain is often performed by estimating a multiplicative mask to extract clean speech. However, most neural network-based methods perform point estimation, i.e., their output consists of a single mask. In this paper, we study the benefits of modeling uncertainty in neural network-based speech enhancement. For this, our neural network is trained to map a noisy spectrogram to the Wiener filter and its associated variance, which quantifies uncertainty, based on the maximum a posteriori (MAP) inference of spectral coefficients. By estimating the distribution instead of the point estimate, one can model the uncertainty associated with each estimate. We further propose to use the estimated Wiener filter and its uncertainty to build an approximate MAP (A-MAP) estimator of spectral magnitudes, which in turn is combined with the MAP inference of spectral coefficients to form a hybrid loss function to jointly reinforce the estimation. Experimental results on different datasets show that the proposed method can not only capture the uncertainty associated with the estimated filters, but also yield a higher enhancement performance over comparable models that do not take uncertainty into account.

Jean-Marie Lemercier, Simon Welker, Timo Gerkmann

We present in this paper an informed single-channel dereverberation method based on conditional generation with diffusion models. With knowledge of the room impulse response, the anechoic utterance is generated via reverse diffusion using a measurement consistency criterion coupled with a neural network that represents the clean speech prior. The proposed approach is largely more robust to measurement noise compared to a state-of-the-art informed single-channel dereverberation method, especially for non-stationary noise. Furthermore, we compare to other blind dereverberation methods using diffusion models and show superiority of the proposed approach for large reverberation times. We motivate the presented algorithm by introducing an extension for blind dereverberation allowing joint estimation of the room impulse response and anechoic speech. Audio samples and code can be found online (https://uhh.de/inf-sp-derev-dps).

Kristina Tesch, Nils-Hendrik Mohrmann, Timo Gerkmann

Employing deep neural networks (DNNs) to directly learn filters for multi-channel speech enhancement has potentially two key advantages over a traditional approach combining a linear spatial filter with an independent tempo-spectral post-filter: 1) non-linear spatial filtering allows to overcome potential restrictions originating from a linear processing model and 2) joint processing of spatial and tempo-spectral information allows to exploit interdependencies between different sources of information. A variety of DNN-based non-linear filters have been proposed recently, for which good enhancement performance is reported. However, little is known about the internal mechanisms which turns network architecture design into a game of chance. Therefore, in this paper, we perform experiments to better understand the internal processing of spatial, spectral and temporal information by DNN-based non-linear filters. On the one hand, our experiments in a difficult speech extraction scenario confirm the importance of non-linear spatial filtering, which outperforms an oracle linear spatial filter by 0.24 POLQA score. On the other hand, we demonstrate that joint processing results in a large performance gap of 0.4 POLQA score between network architectures exploiting spectral versus temporal information besides spatial information.

Jean-Marie Lemercier, Joachim Thiemann, Raphael Koning et al.

This work focuses on online dereverberation for hearing devices using the weighted prediction error (WPE) algorithm. WPE filtering requires an estimate of the target speech power spectral density (PSD). Recently deep neural networks (DNNs) have been used for this task. However, these approaches optimize the PSD estimate which only indirectly affects the WPE output, thus potentially resulting in limited dereverberation. In this paper, we propose an end-to-end approach specialized for online processing, that directly optimizes the dereverberated output signal. In addition, we propose to adapt it to the needs of different types of hearing-device users by modifying the optimization target as well as the WPE algorithm characteristics used in training. We show that the proposed end-to-end approach outperforms the traditional and conventional DNN-supported WPEs on a noise-free version of the WHAMR! dataset.

Danilo de Oliveira, Julius Richter, Jean-Marie Lemercier et al.

Since its inception, the field of deep speech enhancement has been dominated by predictive (discriminative) approaches, such as spectral mapping or masking. Recently, however, novel generative approaches have been applied to speech enhancement, attaining good denoising performance with high subjective quality scores. At the same time, advances in deep learning also allowed for the creation of neural network-based metrics, which have desirable traits such as being able to work without a reference (non-intrusively). Since generatively enhanced speech tends to exhibit radically different residual distortions, its evaluation using instrumental speech metrics may behave differently compared to predictively enhanced speech. In this paper, we evaluate the performance of the same speech enhancement backbone trained under predictive and generative paradigms on a variety of metrics and show that intrusive and non-intrusive measures correlate differently for each paradigm. This analysis motivates the search for metrics that can together paint a complete and unbiased picture of speech enhancement performance, irrespective of the model's training process.

Huajian Fang, Niklas Wittmer, Johannes Twiefel et al.

Human-robot interaction relies on a noise-robust audio processing module capable of estimating target speech from audio recordings impacted by environmental noise, as well as self-induced noise, so-called ego-noise. While external ambient noise sources vary from environment to environment, ego-noise is mainly caused by the internal motors and joints of a robot. Ego-noise and environmental noise reduction are often decoupled, i.e., ego-noise reduction is performed without considering environmental noise. Recently, a variational autoencoder (VAE)-based speech model has been combined with a fully adaptive non-negative matrix factorization (NMF) noise model to recover clean speech under different environmental noise disturbances. However, its enhancement performance is limited in adverse acoustic scenarios involving, e.g. ego-noise. In this paper, we propose a multichannel partially adaptive scheme to jointly model ego-noise and environmental noise utilizing the VAE-NMF framework, where we take advantage of spatially and spectrally structured characteristics of ego-noise by pre-training the ego-noise model, while retaining the ability to adapt to unknown environmental noise. Experimental results show that our proposed approach outperforms the methods based on a completely fixed scheme and a fully adaptive scheme when ego-noise and environmental noise are present simultaneously.

Danilo de Oliveira, Timo Gerkmann

Much research effort is being applied to the task of compressing the knowledge of self-supervised models, which are powerful, yet large and memory consuming. In this work, we show that the original method of knowledge distillation (and its more recently proposed extension, decoupled knowledge distillation) can be applied to the task of distilling HuBERT. In contrast to methods that focus on distilling internal features, this allows for more freedom in the network architecture of the compressed model. We thus propose to distill HuBERT's Transformer layers into an LSTM-based distilled model that reduces the number of parameters even below DistilHuBERT and at the same time shows improved performance in automatic speech recognition.

Navin Raj Prabhu, Nale Lehmann-Willenbrock, Timo Gerkmann

As different people perceive others' emotional expressions differently, their annotation in terms of arousal and valence are per se subjective. To address this, these emotion annotations are typically collected by multiple annotators and averaged across annotators in order to obtain labels for arousal and valence. However, besides the average, also the uncertainty of a label is of interest, and should also be modeled and predicted for automatic emotion recognition. In the literature, for simplicity, label uncertainty modeling is commonly approached with a Gaussian assumption on the collected annotations. However, as the number of annotators is typically rather small due to resource constraints, we argue that the Gaussian approach is a rather crude assumption. In contrast, in this work we propose to model the label distribution using a Student's t-distribution which allows us to account for the number of annotations available. With this model, we derive the corresponding Kullback-Leibler divergence based loss function and use it to train an estimator for the distribution of emotion labels, from which the mean and uncertainty can be inferred. Through qualitative and quantitative analysis, we show the benefits of the t-distribution over a Gaussian distribution. We validate our proposed method on the AVEC'16 dataset. Results reveal that our t-distribution based approach improves over the Gaussian approach with state-of-the-art uncertainty modeling results in speech-based emotion recognition, along with an optimal and even faster convergence.

Jean-Marie Lemercier, Joachim Thiemann, Raphael Koning et al.

A two-stage lightweight online dereverberation algorithm for hearing devices is presented in this paper. The approach combines a multi-channel multi-frame linear filter with a single-channel single-frame post-filter. Both components rely on power spectral density (PSD) estimates provided by deep neural networks (DNNs). By deriving new metrics analyzing the dereverberation performance in various time ranges, we confirm that directly optimizing for a criterion at the output of the multi-channel linear filtering stage results in a more efficient dereverberation as compared to placing the criterion at the output of the DNN to optimize the PSD estimation. More concretely, we show that training this stage end-to-end helps further remove the reverberation in the range accessible to the filter, thus increasing the \textit{early-to-moderate} reverberation ratio. We argue and demonstrate that it can then be well combined with a post-filtering stage to efficiently suppress the residual late reverberation, thereby increasing the \textit{early-to-final} reverberation ratio. This proposed two stage procedure is shown to be both very effective in terms of dereverberation performance and computational demands, as compared to e.g. recent state-of-the-art DNN approaches. Furthermore, the proposed two-stage system can be adapted to the needs of different types of hearing-device users by controlling the amount of reduction of early reflections.

Navin Raj Prabhu, Nale Lehmann-Willenbrock, Timo Gerkmann

Speech emotion conversion aims to convert the expressed emotion of a spoken utterance to a target emotion while preserving the lexical information and the speaker's identity. In this work, we specifically focus on in-the-wild emotion conversion where parallel data does not exist, and the problem of disentangling lexical, speaker, and emotion information arises. In this paper, we introduce a methodology that uses self-supervised networks to disentangle the lexical, speaker, and emotional content of the utterance, and subsequently uses a HiFiGAN vocoder to resynthesise the disentangled representations to a speech signal of the targeted emotion. For better representation and to achieve emotion intensity control, we specifically focus on the aro\-usal dimension of continuous representations, as opposed to performing emotion conversion on categorical representations. We test our methodology on the large in-the-wild MSP-Podcast dataset. Results reveal that the proposed approach is aptly conditioned on the emotional content of input speech and is capable of synthesising natural-sounding speech for a target emotion. Results further reveal that the methodology better synthesises speech for mid-scale arousal (2 to 6) than for extreme arousal (1 and 7).

Huajian Fang, Timo Gerkmann

Single-channel deep speech enhancement approaches often estimate a single multiplicative mask to extract clean speech without a measure of its accuracy. Instead, in this work, we propose to quantify the uncertainty associated with clean speech estimates in neural network-based speech enhancement. Predictive uncertainty is typically categorized into aleatoric uncertainty and epistemic uncertainty. The former accounts for the inherent uncertainty in data and the latter corresponds to the model uncertainty. Aiming for robust clean speech estimation and efficient predictive uncertainty quantification, we propose to integrate statistical complex Gaussian mixture models (CGMMs) into a deep speech enhancement framework. More specifically, we model the dependency between input and output stochastically by means of a conditional probability density and train a neural network to map the noisy input to the full posterior distribution of clean speech, modeled as a mixture of multiple complex Gaussian components. Experimental results on different datasets show that the proposed algorithm effectively captures predictive uncertainty and that combining powerful statistical models and deep learning also delivers a superior speech enhancement performance.

Jean-Marie Lemercier, Julian Tobergte, Timo Gerkmann

In this paper, we present a scheme for extending deep neural network-based multiplicative maskers to deep subband filters for speech restoration in the time-frequency domain. The resulting method can be generically applied to any deep neural network providing masks in the time-frequency domain, while requiring only few more trainable parameters and a computational overhead that is negligible for state-of-the-art neural networks. We demonstrate that the resulting deep subband filtering scheme outperforms multiplicative masking for dereverberation, while leaving the denoising performance virtually the same. We argue that this is because deep subband filtering in the time-frequency domain fits the subband approximation often assumed in the dereverberation literature, whereas multiplicative masking corresponds to the narrowband approximation generally employed for denoising.

Jean-Marie Lemercier, Joachim Thiemann, Raphael Koning et al.

In this paper we present a method for single-channel wind noise reduction using our previously proposed diffusion-based stochastic regeneration model combining predictive and generative modelling. We introduce a non-additive speech in noise model to account for the non-linear deformation of the membrane caused by the wind flow and possible clipping. We show that our stochastic regeneration model outperforms other neural-network-based wind noise reduction methods as well as purely predictive and generative models, on a dataset using simulated and real-recorded wind noise. We further show that the proposed method generalizes well by testing on an unseen dataset with real-recorded wind noise. Audio samples, data generation scripts and code for the proposed methods can be found online (https://uhh.de/inf-sp-storm-wind).

Jean-Marie Lemercier, Joachim Thiemann, Raphael Koning et al.

In this paper, a neural network-augmented algorithm for noise-robust online dereverberation with a Kalman filtering variant of the weighted prediction error (WPE) method is proposed. The filter stochastic variations are predicted by a deep neural network (DNN) trained end-to-end using the filter residual error and signal characteristics. The presented framework allows for robust dereverberation on a single-channel noisy reverberant dataset similar to WHAMR!. The Kalman filtering WPE introduces distortions in the enhanced signal when predicting the filter variations from the residual error only, if the target speech power spectral density is not perfectly known and the observation is noisy. The proposed approach avoids these distortions by correcting the filter variations estimation in a data-driven way, increasing the robustness of the method to noisy scenarios. Furthermore, it yields a strong dereverberation and denoising performance compared to a DNN-supported recursive least squares variant of WPE, especially for highly noisy inputs.

Bunlong Lay, Sebastian Zaczek, Kristina Tesch et al.

Single-channel speech separation is a crucial task for enhancing speech recognition systems in multi-speaker environments. This paper investigates the robustness of state-of-the-art Neural Network models in scenarios where the pitch differences between speakers are minimal. Building on earlier findings by Ditter and Gerkmann, which identified a significant performance drop for the 2018 Chimera++ under similar-pitch conditions, our study extends the analysis to more recent and sophisticated Neural Network models. Our experiments reveal that modern models have substantially reduced the performance gap for matched training and testing conditions. However, a substantial performance gap persists under mismatched conditions, with models performing well for large pitch differences but showing worse performance if the speakers' pitches are similar. These findings motivate further research into the generalizability of speech separation models to similar-pitch speakers and unseen data.

Simon Welker, Bunlong Lay, Maris Hillemann et al.

Diffusion-based generative models have greatly impacted the speech processing field in recent years, exhibiting high speech naturalness and spawning a new research direction. Their application in real-time communication is, however, still lagging behind due to their computation-heavy nature involving multiple calls of large DNNs. Here, we present Stream.FM, a frame-causal flow-based generative model with an algorithmic latency of 32 milliseconds (ms) and a total latency of 48 ms, paving the way for generative speech processing in real-time communication. We propose a buffered streaming inference scheme and an optimized DNN architecture, show how learned few-step numerical solvers can boost output quality at a fixed compute budget, explore model weight compression to find favorable points along a compute/quality tradeoff, and contribute a model variant with 24 ms total latency for the speech enhancement task. Our work looks beyond theoretical latencies, showing that high-quality streaming generative speech processing can be realized on consumer GPUs available today. Stream.FM can solve a variety of speech processing tasks in a streaming fashion: speech enhancement, dereverberation, codec post-filtering, bandwidth extension, STFT phase retrieval, and Mel vocoding. As we verify through comprehensive evaluations and a MUSHRA listening test, Stream.FM establishes a state-of-the-art for generative streaming speech restoration, exhibits only a reasonable reduction in quality compared to a non-streaming variant, and outperforms our recent work (Diffusion Buffer) on generative streaming speech enhancement while operating at a lower latency.

Jean-Marie Lemercier, Julius Richter, Simon Welker et al.

With the development of audio playback devices and fast data transmission, the demand for high sound quality is rising for both entertainment and communications. In this quest for better sound quality, challenges emerge from distortions and interferences originating at the recording side or caused by an imperfect transmission pipeline. To address this problem, audio restoration methods aim to recover clean sound signals from the corrupted input data. We present here audio restoration algorithms based on diffusion models, with a focus on speech enhancement and music restoration tasks. Traditional approaches, often grounded in handcrafted rules and statistical heuristics, have shaped our understanding of audio signals. In the past decades, there has been a notable shift towards data-driven methods that exploit the modeling capabilities of DNNs. Deep generative models, and among them diffusion models, have emerged as powerful techniques for learning complex data distributions. However, relying solely on DNN-based learning approaches carries the risk of reducing interpretability, particularly when employing end-to-end models. Nonetheless, data-driven approaches allow more flexibility in comparison to statistical model-based frameworks, whose performance depends on distributional and statistical assumptions that can be difficult to guarantee. Here, we aim to show that diffusion models can combine the best of both worlds and offer the opportunity to design audio restoration algorithms with a good degree of interpretability and a remarkable performance in terms of sound quality. We explain the diffusion formalism and its application to the conditional generation of clean audio signals. We believe that diffusion models open an exciting field of research with the potential to spawn new audio restoration algorithms that are natural-sounding and remain robust in difficult acoustic situations.

Jakob Kienegger, Timo Gerkmann

Deep spatially selective filters achieve high-quality enhancement with real-time capable architectures for stationary speakers of known directions. To retain this level of performance in dynamic scenarios when only the speakers' initial directions are given, accurate, yet computationally lightweight tracking algorithms become necessary. Assuming a frame-wise causal processing style, temporal feedback allows for leveraging the enhanced speech signal to improve tracking performance. In this work, we investigate strategies to incorporate the enhanced signal into lightweight tracking algorithms and autoregressively guide deep spatial filters. Our proposed Bayesian tracking algorithms are compatible with arbitrary deep spatial filters. To increase the realism of simulated trajectories during development and evaluation, we propose and publish a novel dataset based on the social force model. Results validate that the autoregressive incorporation significantly improves the accuracy of our Bayesian trackers, resulting in superior enhancement with none or only negligibly increased computational overhead. Real-world recordings complement these findings and demonstrate the generalizability of our methods to unseen, challenging acoustic conditions.

Eloi Moliner, Jean-Marie Lemercier, Simon Welker et al.

In this paper, we present an unsupervised single-channel method for joint blind dereverberation and room impulse response estimation, based on posterior sampling with diffusion models. We parameterize the reverberation operator using a filter with exponential decay for each frequency subband, and iteratively estimate the corresponding parameters as the speech utterance gets refined along the reverse diffusion trajectory. A measurement consistency criterion enforces the fidelity of the generated speech with the reverberant measurement, while an unconditional diffusion model implements a strong prior for clean speech generation. Without any knowledge of the room impulse response nor any coupled reverberant-anechoic data, we can successfully perform dereverberation in various acoustic scenarios. Our method significantly outperforms previous blind unsupervised baselines, and we demonstrate its increased robustness to unseen acoustic conditions in comparison to blind supervised methods. Audio samples and code are available online.

Simon Welker, Matthew Le, Ricky T. Q. Chen et al.

We propose FlowDec, a neural full-band audio codec for general audio sampled at 48 kHz that combines non-adversarial codec training with a stochastic postfilter based on a novel conditional flow matching method. Compared to the prior work ScoreDec which is based on score matching, we generalize from speech to general audio and move from 24 kbit/s to as low as 4 kbit/s, while improving output quality and reducing the required postfilter DNN evaluations from 60 to 6 without any fine-tuning or distillation techniques. We provide theoretical insights and geometric intuitions for our approach in comparison to ScoreDec as well as another recent work that uses flow matching, and conduct ablation studies on our proposed components. We show that FlowDec is a competitive alternative to the recent GAN-dominated stream of neural codecs, achieving FAD scores better than those of the established GAN-based codec DAC and listening test scores that are on par, and producing qualitatively more natural reconstructions for speech and harmonic structures in music.

Bunlong Lay, Timo Gerkmann

Diffusion models proved to be powerful models for generative speech enhancement. In recent SGMSE+ approaches, training involves a stochastic differential equation for the diffusion process, adding both Gaussian and environmental noise to the clean speech signal gradually. The speech enhancement performance varies depending on the choice of the stochastic differential equation that controls the evolution of the mean and the variance along the diffusion processes when adding environmental and Gaussian noise. In this work, we highlight that the scale of the variance is a dominant parameter for speech enhancement performance and show that it controls the tradeoff between noise attenuation and speech distortions. More concretely, we show that a larger variance increases the noise attenuation and allows for reducing the computational footprint, as fewer function evaluations for generating the estimate are required

Bunlong Lay, Rostislav Makarov, Timo Gerkmann

Diffusion models are a class of generative models that have been recently used for speech enhancement with remarkable success but are computationally expensive at inference time. Therefore, these models are impractical for processing streaming data in real-time. In this work, we adapt a sliding window diffusion framework to the speech enhancement task. Our approach progressively corrupts speech signals through time, assigning more noise to frames close to the present in a buffer. This approach outputs denoised frames with a delay proportional to the chosen buffer size, enabling a trade-off between performance and latency. Empirical results demonstrate that our method outperforms standard diffusion models and runs efficiently on a GPU, achieving an input-output latency in the order of 0.3 to 1 seconds. This marks the first practical diffusion-based solution for online speech enhancement.

Danilo de Oliveira, Julius Richter, Jean-Marie Lemercier et al.

Diffusion models have found great success in generating high quality, natural samples of speech, but their potential for density estimation for speech has so far remained largely unexplored. In this work, we leverage an unconditional diffusion model trained only on clean speech for the assessment of speech quality. We show that the quality of a speech utterance can be assessed by estimating the likelihood of a corresponding sample in the terminating Gaussian distribution, obtained via a deterministic noising process. The resulting method is purely unsupervised, trained only on clean speech, and therefore does not rely on annotations. Our diffusion-based approach leverages clean speech priors to assess quality based on how the input relates to the learned distribution of clean data. Our proposed log-likelihoods show promising results, correlating well with intrusive speech quality metrics and showing the best correlation with human scores in a listening experiment.

Jakob Kienegger, Timo Gerkmann

Recent speaker extraction methods using deep non-linear spatial filtering perform exceptionally well when the target direction is known and stationary. However, spatially dynamic scenarios are considerably more challenging due to time-varying spatial features and arising ambiguities, e.g. when moving speakers cross. While in a static scenario it may be easy for a user to point to the target's direction, manually tracking a moving speaker is impractical. Instead of relying on accurate time-dependent directional cues, which we refer to as strong guidance, in this paper we propose a weakly guided extraction method solely depending on the target's initial position to cope with spatial dynamic scenarios. By incorporating our own deep tracking algorithm and developing a joint training strategy on a synthetic dataset, we demonstrate the proficiency of our approach in resolving spatial ambiguities and even outperform a mismatched, but strongly guided extraction method.

Bunlong Lay, Rostislav Makarov, Simon Welker et al.

Online Speech Enhancement was mainly reserved for predictive models. A key advantage of these models is that for an incoming signal frame from a stream of data, the model is called only once for enhancement. In contrast, generative Speech Enhancement models often require multiple calls, resulting in a computational complexity that is too high for many online speech enhancement applications. This work presents the Diffusion Buffer, a generative diffusion-based Speech Enhancement model which only requires one neural network call per incoming signal frame from a stream of data and performs enhancement in an online fashion on a consumer-grade GPU. The key idea of the Diffusion Buffer is to align physical time with Diffusion time-steps. The approach progressively denoises frames through physical time, where past frames have more noise removed. Consequently, an enhanced frame is output to the listener with a delay defined by the Diffusion Buffer, and the output frame has a corresponding look-ahead. In this work, we extend upon our previous work by carefully designing a 2D convolutional UNet architecture that specifically aligns with the Diffusion Buffer's look-ahead. We observe that the proposed UNet improves performance, particularly when the algorithmic latency is low. Moreover, we show that using a Data Prediction loss instead of Denoising Score Matching loss enables flexible control over the trade-off between algorithmic latency and quality during inference. The extended Diffusion Buffer equipped with a novel NN and loss function drastically reduces the algorithmic latency from 320 - 960 ms to 32 - 176 ms with an even increased performance. While it has been shown before that offline generative diffusion models outperform predictive approaches in unseen noisy speech data, we confirm that the online Diffusion Buffer also outperforms its predictive counterpart on unseen noisy speech data.

Simon Welker, Lorenz Kuger, Tim Roith et al.

In this work, we present and investigate the novel blind inverse problem of position-blind ptychography, i.e., ptychographic phase retrieval without any knowledge of scan positions, which then must be recovered jointly with the image. The motivation for this problem comes from single-particle diffractive X-ray imaging, where particles in random orientations are illuminated and a set of diffraction patterns is collected. If one uses a highly focused X-ray beam, the measurements would also become sensitive to the beam positions relative to each particle and therefore ptychographic, but these positions are also unknown. We investigate the viability of image reconstruction in a simulated, simplified 2-D variant of this difficult problem, using variational inference with modern data-driven image priors in the form of score-based diffusion models. We find that, with the right illumination structure and a strong prior, one can achieve reliable and successful image reconstructions even under measurement noise, in all except the most difficult evaluated imaging scenario.

Rostislav Makarov, Lea Schönherr, Timo Gerkmann

Machine learning approaches for speech enhancement are becoming increasingly expressive, enabling ever more powerful modifications of input signals. In this paper, we demonstrate that this expressiveness introduces a vulnerability: advanced speech enhancement models can be susceptible to adversarial attacks. Specifically, we show that adversarial noise, carefully crafted and psychoacoustically masked by the original input, can be injected such that the enhanced speech output conveys an entirely different semantic meaning. We experimentally verify that contemporary predictive speech enhancement models can indeed be manipulated in this way. Furthermore, we highlight that diffusion models with stochastic samplers exhibit inherent robustness to such adversarial attacks by design.

Simon Welker, Tal Peer, Timo Gerkmann

The task of Mel vocoding, i.e., the inversion of a Mel magnitude spectrogram to an audio waveform, is still a key component in many text-to-speech (TTS) systems today. Based on generative flow matching, our prior work on generative STFT phase retrieval (DiffPhase), and the pseudoinverse operator of the Mel filterbank, we develop MelFlow, a streaming-capable generative Mel vocoder for speech sampled at 16 kHz with an algorithmic latency of only 32 ms and a total latency of 48 ms. We show real-time streaming capability at this latency not only in theory, but in practice on a consumer laptop GPU. Furthermore, we show that our model achieves substantially better PESQ and SI-SDR values compared to well-established not streaming-capable baselines for Mel vocoding including HiFi-GAN.

Jakob Kienegger, Alina Mannanova, Huajian Fang et al.

Recent works on deep non-linear spatially selective filters demonstrate exceptional enhancement performance with computationally lightweight architectures for stationary speakers of known directions. However, to maintain this performance in dynamic scenarios, resource-intensive data-driven tracking algorithms become necessary to provide precise spatial guidance conditioned on the initial direction of a target speaker. As this additional computational overhead hinders application in resource-constrained scenarios such as real-time speech enhancement, we present a novel strategy utilizing a low-complexity tracking algorithm in the form of a particle filter instead. Assuming a causal, sequential processing style, we introduce temporal feedback to leverage the enhanced speech signal of the spatially selective filter to compensate for the limited modeling capabilities of the particle filter. Evaluation on a synthetic dataset illustrates how the autoregressive interplay between both algorithms drastically improves tracking accuracy and leads to strong enhancement performance. A listening test with real-world recordings complements these findings by indicating a clear trend towards our proposed self-steering pipeline as preferred choice over comparable methods.

Jakob Kienegger, Alina Mannanova, Timo Gerkmann

Due to their robustness and flexibility, neural-driven beamformers are a popular choice for speech separation in challenging environments with a varying amount of simultaneous speakers alongside noise and reverberation. Time-frequency masks and relative directions of the speakers regarding a fixed spatial grid can be used to estimate the beamformer's parameters. To some degree, speaker-independence is achieved by ensuring a greater amount of spatial partitions than speech sources. In this work, we analyze how to encode both mask and positioning into such a grid to enable joint estimation of both quantities. We propose mask-weighted spatial likelihood coding and show that it achieves considerable performance in both tasks compared to baseline encodings optimized for either localization or mask estimation. In the same setup, we demonstrate superiority for joint estimation of both quantities. Conclusively, we propose a universal approach which can replace an upstream sound source localization system solely by adapting the training framework, making it highly relevant in performance-critical scenarios.

Julius Richter, Yi-Chiao Wu, Steven Krenn et al.

We release the EARS (Expressive Anechoic Recordings of Speech) dataset, a high-quality speech dataset comprising 107 speakers from diverse backgrounds, totaling in 100 hours of clean, anechoic speech data. The dataset covers a large range of different speaking styles, including emotional speech, different reading styles, non-verbal sounds, and conversational freeform speech. We benchmark various methods for speech enhancement and dereverberation on the dataset and evaluate their performance through a set of instrumental metrics. In addition, we conduct a listening test with 20 participants for the speech enhancement task, where a generative method is preferred. We introduce a blind test set that allows for automatic online evaluation of uploaded data. Dataset download links and automatic evaluation server can be found online.

Danilo de Oliveira, Simon Welker, Julius Richter et al.

To obtain improved speech enhancement models, researchers often focus on increasing performance according to specific instrumental metrics. However, when the same metric is used in a loss function to optimize models, it may be detrimental to aspects that the given metric does not see. The goal of this paper is to illustrate the risk of overfitting a speech enhancement model to the metric used for evaluation. For this, we introduce enhancement models that exploit the widely used PESQ measure. Our "PESQetarian" model achieves 3.82 PESQ on VB-DMD while scoring very poorly in a listening experiment. While the obtained PESQ value of 3.82 would imply "state-of-the-art" PESQ-performance on the VB-DMD benchmark, our examples show that when optimizing w.r.t. a metric, an isolated evaluation on the same metric may be misleading. Instead, other metrics should be included in the evaluation and the resulting performance predictions should be confirmed by listening.

Héctor Martel, Julius Richter, Kai Li et al.

We propose Audio-Visual Lightweight ITerative model (AVLIT), an effective and lightweight neural network that uses Progressive Learning (PL) to perform audio-visual speech separation in noisy environments. To this end, we adopt the Asynchronous Fully Recurrent Convolutional Neural Network (A-FRCNN), which has shown successful results in audio-only speech separation. Our architecture consists of an audio branch and a video branch, with iterative A-FRCNN blocks sharing weights for each modality. We evaluated our model in a controlled environment using the NTCD-TIMIT dataset and in-the-wild using a synthetic dataset that combines LRS3 and WHAM!. The experiments demonstrate the superiority of our model in both settings with respect to various audio-only and audio-visual baselines. Furthermore, the reduced footprint of our model makes it suitable for low resource applications.

Danilo de Oliveira, Navin Raj Prabhu, Timo Gerkmann

In large part due to their implicit semantic modeling, self-supervised learning (SSL) methods have significantly increased the performance of valence recognition in speech emotion recognition (SER) systems. Yet, their large size may often hinder practical implementations. In this work, we take HuBERT as an example of an SSL model and analyze the relevance of each of its layers for SER. We show that shallow layers are more important for arousal recognition while deeper layers are more important for valence. This observation motivates the importance of additional textual information for accurate valence recognition, as the distilled framework lacks the depth of its large-scale SSL teacher. Thus, we propose an audio-textual distilled SSL framework that, while having only ~20% of the trainable parameters of a large SSL model, achieves on par performance across the three emotion dimensions (arousal, valence, dominance) on the MSP-Podcast v1.10 dataset.

Huajian Fang, Dennis Becker, Stefan Wermter et al.

Supervised masking approaches in the time-frequency domain aim to employ deep neural networks to estimate a multiplicative mask to extract clean speech. This leads to a single estimate for each input without any guarantees or measures of reliability. In this paper, we study the benefits of modeling uncertainty in clean speech estimation. Prediction uncertainty is typically categorized into aleatoric uncertainty and epistemic uncertainty. The former refers to inherent randomness in data, while the latter describes uncertainty in the model parameters. In this work, we propose a framework to jointly model aleatoric and epistemic uncertainties in neural network-based speech enhancement. The proposed approach captures aleatoric uncertainty by estimating the statistical moments of the speech posterior distribution and explicitly incorporates the uncertainty estimate to further improve clean speech estimation. For epistemic uncertainty, we investigate two Bayesian deep learning approaches: Monte Carlo dropout and Deep ensembles to quantify the uncertainty of the neural network parameters. Our analyses show that the proposed framework promotes capturing practical and reliable uncertainty, while combining different sources of uncertainties yields more reliable predictive uncertainty estimates. Furthermore, we demonstrate the benefits of modeling uncertainty on speech enhancement performance by evaluating the framework on different datasets, exhibiting notable improvement over comparable models that fail to account for uncertainty.

Simon Welker, Julius Richter, Timo Gerkmann

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.

Tal Peer, Timo Gerkmann

Algorithmic latency in speech processing is dominated by the frame length used for Fourier analysis, which in turn limits the achievable performance of magnitude-centric approaches. As previous studies suggest the importance of phase grows with decreasing frame length, this work presents a systematical study on the contribution of phase and magnitude in modern Deep Neural Network (DNN)-based speech enhancement at different frame lengths. Results indicate that DNNs can successfully estimate phase when using short frames, with similar or better overall performance compared to using longer frames. Thus, interestingly, modern phase-aware DNNs allow for low-latency speech enhancement at high quality.