Zheng-Hua Tan

Sarthak Yadav, Sergios Theodoridis, Lars Kai Hansen et al.

In this work, we propose a Multi-Window Masked Autoencoder (MW-MAE) fitted with a novel Multi-Window Multi-Head Attention (MW-MHA) module that facilitates the modelling of local-global interactions in every decoder transformer block through attention heads of several distinct local and global windows. Empirical results on ten downstream audio tasks show that MW-MAEs consistently outperform standard MAEs in overall performance and learn better general-purpose audio representations, along with demonstrating considerably better scaling characteristics. Investigating attention distances and entropies reveals that MW-MAE encoders learn heads with broader local and global attention. Analyzing attention head feature representations through Projection Weighted Canonical Correlation Analysis (PWCCA) shows that attention heads with the same window sizes across the decoder layers of the MW-MAE learn correlated feature representations which enables each block to independently capture local and global information, leading to a decoupled decoder feature hierarchy. Code for feature extraction and downstream experiments along with pre-trained models will be released publically.

Cristian J. Vaca-Rubio, Dariush Salami, Petar Popovski et al.

Since electromagnetic signals are omnipresent, Radio Frequency (RF)-sensing has the potential to become a universal sensing mechanism with applications in localization, smart-home, retail, gesture recognition, intrusion detection, etc. Two emerging technologies in RF-sensing, namely sensing through Large Intelligent Surfaces (LISs) and mmWave Frequency-Modulated Continuous-Wave (FMCW) radars, have been successfully applied to a wide range of applications. In this work, we compare LIS and mmWave radars for localization in real-world and simulated environments. In our experiments, the mmWave radar achieves 0.71 Intersection Over Union (IOU) and 3cm error for bounding boxes, while LIS has 0.56 IOU and 10cm distance error. Although the radar outperforms the LIS in terms of accuracy, LIS features additional applications in communication in addition to sensing scenarios.

Cristian J. Vaca-Rubio, Roberto Pereira, Xavier Mestre et al.

Environmental scene reconstruction is of great interest for autonomous robotic applications, since an accurate representation of the environment is necessary to ensure safe interaction with robots. Equally important, it is also vital to ensure reliable communication between the robot and its controller. Large Intelligent Surface (LIS) is a technology that has been extensively studied due to its communication capabilities. Moreover, due to the number of antenna elements, these surfaces arise as a powerful solution to radio sensing. This paper presents a novel method to translate radio environmental maps obtained at the LIS to floor plans of the indoor environment built of scatterers spread along its area. The usage of a Least Squares (LS) based method, U-Net (UN) and conditional Generative Adversarial Networks (cGANs) were leveraged to perform this task. We show that the floor plan can be correctly reconstructed using both local and global measurements.

Iván López-Espejo, Ram C. M. C. Shekar, Zheng-Hua Tan et al.

In the context of keyword spotting (KWS), the replacement of handcrafted speech features by learnable features has not yielded superior KWS performance. In this study, we demonstrate that filterbank learning outperforms handcrafted speech features for KWS whenever the number of filterbank channels is severely decreased. Reducing the number of channels might yield certain KWS performance drop, but also a substantial energy consumption reduction, which is key when deploying common always-on KWS on low-resource devices. Experimental results on a noisy version of the Google Speech Commands Dataset show that filterbank learning adapts to noise characteristics to provide a higher degree of robustness to noise, especially when dropout is integrated. Thus, switching from typically used 40-channel log-Mel features to 8-channel learned features leads to a relative KWS accuracy loss of only 3.5% while simultaneously achieving a 6.3x energy consumption reduction.

Kevin Wilkinghoff, Keisuke Imoto, Zheng-Hua Tan

Anomalous sound detection (ASD) benchmarks typically assume that the identity of the monitored machine is known at test time and that recordings are evaluated in a machine-wise manner. However, in realistic monitoring scenarios with multiple known machines operating concurrently, test recordings may not be reliably attributable to a specific machine, and requiring machine identity imposes deployment constraints such as dedicated sensors per machine. To reveal performance degradations and method-specific differences in robustness that are hidden under standard machine-wise evaluation, we consider a minimal modification of the ASD evaluation protocol in which test recordings from multiple machines are merged and evaluated jointly without access to machine identity at inference time. Training data and evaluation metrics remain unchanged, and machine identity labels are used only for post hoc evaluation. Experiments with representative ASD methods show that relaxing this assumption reveals performance degradations and method-specific differences in robustness that are hidden under standard machine-wise evaluation, and that these degradations are strongly related to implicit machine identification accuracy.

Juan F. Montesinos, Daniel Michelsanti, Gloria Haro et al.

Audio and visual modalities are inherently connected in speech signals: lip movements and facial expressions are correlated with speech sounds. This motivates studies that incorporate the visual modality to enhance an acoustic speech signal or even restore missing audio information. Specifically, this paper focuses on the problem of audio-visual speech inpainting, which is the task of synthesizing the speech in a corrupted audio segment in a way that it is consistent with the corresponding visual content and the uncorrupted audio context. We present an audio-visual transformer-based deep learning model that leverages visual cues that provide information about the content of the corrupted audio. It outperforms the previous state-of-the-art audio-visual model and audio-only baselines. We also show how visual features extracted with AV-HuBERT, a large audio-visual transformer for speech recognition, are suitable for synthesizing speech.

Holger Severin Bovbjerg, Zheng-Hua Tan

Keyword Spotting (KWS) models are becoming increasingly integrated into various systems, e.g. voice assistants. To achieve satisfactory performance, these models typically rely on a large amount of labelled data, limiting their applications only to situations where such data is available. Self-supervised Learning (SSL) methods can mitigate such a reliance by leveraging readily-available unlabelled data. Most SSL methods for speech have primarily been studied for large models, whereas this is not ideal, as compact KWS models are generally required. This paper explores the effectiveness of SSL on small models for KWS and establishes that SSL can enhance the performance of small KWS models when labelled data is scarce. We pretrain three compact transformer-based KWS models using Data2Vec, and fine-tune them on a label-deficient setup of the Google Speech Commands data set. It is found that Data2Vec pretraining leads to a significant increase in accuracy, with label-deficient scenarios showing an improvement of 8.22% 11.18% absolute accuracy.

Deividas Eringis, John Leth, Zheng-Hua Tan et al.

In this paper we derive a Probably Approxilmately Correct(PAC)-Bayesian error bound for linear time-invariant (LTI) stochastic dynamical systems with inputs. Such bounds are widespread in machine learning, and they are useful for characterizing the predictive power of models learned from finitely many data points. In particular, with the bound derived in this paper relates future average prediction errors with the prediction error generated by the model on the data used for learning. In turn, this allows us to provide finite-sample error bounds for a wide class of learning/system identification algorithms. Furthermore, as LTI systems are a sub-class of recurrent neural networks (RNNs), these error bounds could be a first step towards PAC-Bayesian bounds for RNNs.

Yuying Xie, Thomas Arildsen, Zheng-Hua Tan

Leveraging the fact that speaker identity and content vary on different time scales, \acrlong{fhvae} (\acrshort{fhvae}) uses different latent variables to symbolize these two attributes. Disentanglement of these attributes is carried out by different prior settings of the corresponding latent variables. For the prior of speaker identity variable, \acrshort{fhvae} assumes it is a Gaussian distribution with an utterance-scale varying mean and a fixed variance. By setting a small fixed variance, the training process promotes identity variables within one utterance gathering close to the mean of their prior. However, this constraint is relatively weak, as the mean of the prior changes between utterances. Therefore, we introduce contrastive learning into the \acrshort{fhvae} framework, to make the speaker identity variables gathering when representing the same speaker, while distancing themselves as far as possible from those of other speakers. The model structure has not been changed in this work but only the training process, thus no additional cost is needed during testing. Voice conversion has been chosen as the application in this paper. Latent variable evaluations include speaker verification and identification for the speaker identity variable, and speech recognition for the content variable. Furthermore, assessments of voice conversion performance are on the grounds of fake speech detection experiments. Results show that the proposed method improves both speaker identity and content feature extraction compared to \acrshort{fhvae}, and has better performance than baseline on conversion.

Christian Heider Nielsen, Zheng-Hua Tan

In recent years, significant progress has been made in deep model-based automatic speech recognition (ASR), leading to its widespread deployment in the real world. At the same time, adversarial attacks against deep ASR systems are highly successful. Various methods have been proposed to defend ASR systems from these attacks. However, existing classification based methods focus on the design of deep learning models while lacking exploration of domain specific features. This work leverages filter bank-based features to better capture the characteristics of attacks for improved detection. Furthermore, the paper analyses the potentials of using speech and non-speech parts separately in detecting adversarial attacks. In the end, considering adverse environments where ASR systems may be deployed, we study the impact of acoustic noise of various types and signal-to-noise ratios. Extensive experiments show that the inverse filter bank features generally perform better in both clean and noisy environments, the detection is effective using either speech or non-speech part, and the acoustic noise can largely degrade the detection performance.

Ioannis Stylianou, Achintya kr. Sarkar, Nauman Dawalatabad et al.

Robust Voice Activity Detection (VAD) remains a challenging task, especially under noisy, diverse, and unseen acoustic conditions. Beyond algorithmic development, a key limitation in advancing VAD research is the lack of large-scale, systematically controlled, and publicly available datasets. To address this, we introduce LibriVAD - a scalable open-source dataset derived from LibriSpeech and augmented with diverse real-world and synthetic noise sources. LibriVAD enables systematic control over speech-to-noise ratio, silence-to-speech ratio (SSR), and noise diversity, and is released in three sizes (15 GB, 150 GB, and 1.5 TB) with two variants (LibriVAD-NonConcat and LibriVAD-Concat) to support different experimental setups. We benchmark multiple feature-model combinations, including waveform, Mel-Frequency Cepstral Coefficients (MFCC), and Gammatone filter bank cepstral coefficients, and introduce the Vision Transformer (ViT) architecture for VAD. Our experiments show that ViT with MFCC features consistently outperforms established VAD models such as boosted deep neural network and convolutional long short-term memory deep neural network across seen, unseen, and out-of-distribution (OOD) conditions, including evaluation on the real-world VOiCES dataset. We further analyze the impact of dataset size and SSR on model generalization, experimentally showing that scaling up dataset size and balancing SSR noticeably and consistently enhance VAD performance under OOD conditions. All datasets, trained models, and code are publicly released to foster reproducibility and accelerate progress in VAD research.

Deividas Eringis, John Leth, Zheng-Hua Tan et al.

In this paper we derive a PAC-Bayesian-Like error bound for a class of stochastic dynamical systems with inputs, namely, for linear time-invariant stochastic state-space models (stochastic LTI systems for short). This class of systems is widely used in control engineering and econometrics, in particular, they represent a special case of recurrent neural networks. In this paper we 1) formalize the learning problem for stochastic LTI systems with inputs, 2) derive a PAC-Bayesian-Like error bound for such systems, 3) discuss various consequences of this error bound.

Kevin Wilkinghoff, Neelu Madan, Juan Miguel Valverde et al.

Anomaly detection aims to identify observations that deviate from expected behavior. Because anomalous events are inherently sparse, most frameworks are trained exclusively on normal data to learn a single reference model of normality. This implicitly assumes that normal behavior can be captured by a single, unconditional reference distribution. In practice, however, anomalies are often context-dependent: A specific observation may be normal under one operating condition, yet anomalous under another. As machine learning systems are deployed in dynamic and heterogeneous environments, these fixed-context assumptions introduce structural ambiguity, i.e., the inability to distinguish contextual variation from genuine abnormality under marginal modeling, leading to unstable performance and unreliable anomaly assessments. While modern sensing systems frequently collect multimodal data capturing complementary aspects of both system behavior and operating conditions, existing methods treat all data streams equally, without distinguishing contextual information from anomaly-relevant signals. As a result, abnormality is often evaluated without explicitly conditioning on operating conditions. We argue that multimodal anomaly detection should be reframed as a cross-modal contextual inference problem, in which modalities play asymmetric roles, separating context from observation, to define abnormality conditionally rather than relative to a single global reference. This perspective has implications for model design, evaluation protocols, and benchmark construction, and outline open research challenges toward robust, context-aware multimodal anomaly detection.

Ioannis Stylianou, Jon Francombe, Pablo Martinez-Nuevo et al.

Conventional audio equalization is a static process that requires manual and cumbersome adjustments to adapt to changing listening contexts (e.g., mood, location, or social setting). In this paper, we introduce a Large Language Model (LLM)-based alternative that maps natural language text prompts to equalization settings. This enables a conversational approach to sound system control. By utilizing data collected from a controlled listening experiment, our models exploit in-context learning and parameter-efficient fine-tuning techniques to reliably align with population-preferred equalization settings. Our evaluation methods, which leverage distributional metrics that capture users' varied preferences, show statistically significant improvements in distributional alignment over random sampling and static preset baselines. These results indicate that LLMs could function as "artificial equalizers," contributing to the development of more accessible, context-aware, and expert-level audio tuning methods.

Wei Rao, Yihui Fu, Yanxin Hu et al.

The ConferencingSpeech 2021 challenge is proposed to stimulate research on far-field multi-channel speech enhancement for video conferencing. The challenge consists of two separate tasks: 1) Task 1 is multi-channel speech enhancement with single microphone array and focusing on practical application with real-time requirement and 2) Task 2 is multi-channel speech enhancement with multiple distributed microphone arrays, which is a non-real-time track and does not have any constraints so that participants could explore any algorithms to obtain high speech quality. Targeting the real video conferencing room application, the challenge database was recorded from real speakers and all recording facilities were located by following the real setup of conferencing room. In this challenge, we open-sourced the list of open source clean speech and noise datasets, simulation scripts, and a baseline system for participants to develop their own system. The final ranking of the challenge will be decided by the subjective evaluation which is performed using Absolute Category Ratings (ACR) to estimate Mean Opinion Score (MOS), speech MOS (S-MOS), and noise MOS (N-MOS). This paper describes the challenge, tasks, datasets, and subjective evaluation. The baseline system which is a complex ratio mask based neural network and its experimental results are also presented.

Jiyang Xie, Zhanyu Ma, and Jianjun Lei et al.

Due to lack of data, overfitting ubiquitously exists in real-world applications of deep neural networks (DNNs). We propose advanced dropout, a model-free methodology, to mitigate overfitting and improve the performance of DNNs. The advanced dropout technique applies a model-free and easily implemented distribution with parametric prior, and adaptively adjusts dropout rate. Specifically, the distribution parameters are optimized by stochastic gradient variational Bayes in order to carry out an end-to-end training. We evaluate the effectiveness of the advanced dropout against nine dropout techniques on seven computer vision datasets (five small-scale datasets and two large-scale datasets) with various base models. The advanced dropout outperforms all the referred techniques on all the datasets.We further compare the effectiveness ratios and find that advanced dropout achieves the highest one on most cases. Next, we conduct a set of analysis of dropout rate characteristics, including convergence of the adaptive dropout rate, the learned distributions of dropout masks, and a comparison with dropout rate generation without an explicit distribution. In addition, the ability of overfitting prevention is evaluated and confirmed. Finally, we extend the application of the advanced dropout to uncertainty inference, network pruning, text classification, and regression. The proposed advanced dropout is also superior to the corresponding referred methods. Codes are available at https://github.com/PRIS-CV/AdvancedDropout.

Xiaoxu Li, Dongliang Chang, Zhanyu Ma et al.

A deep neural network of multiple nonlinear layers forms a large function space, which can easily lead to overfitting when it encounters small-sample data. To mitigate overfitting in small-sample classification, learning more discriminative features from small-sample data is becoming a new trend. To this end, this paper aims to find a subspace of neural networks that can facilitate a large decision margin. Specifically, we propose the Orthogonal Softmax Layer (OSL), which makes the weight vectors in the classification layer remain orthogonal during both the training and test processes. The Rademacher complexity of a network using the OSL is only $\frac{1}{K}$, where $K$ is the number of classes, of that of a network using the fully connected classification layer, leading to a tighter generalization error bound. Experimental results demonstrate that the proposed OSL has better performance than the methods used for comparison on four small-sample benchmark datasets, as well as its applicability to large-sample datasets. Codes are available at: https://github.com/dongliangchang/OSLNet.

Philippe Gonzalez, Zheng-Hua Tan, Jan Østergaard et al.

Diffusion models are a new class of generative models that have shown outstanding performance in image generation literature. As a consequence, studies have attempted to apply diffusion models to other tasks, such as speech enhancement. A popular approach in adapting diffusion models to speech enhancement consists in modelling a progressive transformation between the clean and noisy speech signals. However, one popular diffusion model framework previously laid in image generation literature did not account for such a transformation towards the system input, which prevents from relating the existing diffusion-based speech enhancement systems with the aforementioned diffusion model framework. To address this, we extend this framework to account for the progressive transformation between the clean and noisy speech signals. This allows us to apply recent developments from image generation literature, and to systematically investigate design aspects of diffusion models that remain largely unexplored for speech enhancement, such as the neural network preconditioning, the training loss weighting, the stochastic differential equation (SDE), or the amount of stochasticity injected in the reverse process. We show that the performance of previous diffusion-based speech enhancement systems cannot be attributed to the progressive transformation between the clean and noisy speech signals. Moreover, we show that a proper choice of preconditioning, training loss weighting, SDE and sampler allows to outperform a popular diffusion-based speech enhancement system while using fewer sampling steps, thus reducing the computational cost by a factor of four.

Philippe Gonzalez, Zheng-Hua Tan, Jan Østergaard et al.

Diffusion models are a new class of generative models that have recently been applied to speech enhancement successfully. Previous works have demonstrated their superior performance in mismatched conditions compared to state-of-the art discriminative models. However, this was investigated with a single database for training and another one for testing, which makes the results highly dependent on the particular databases. Moreover, recent developments from the image generation literature remain largely unexplored for speech enhancement. These include several design aspects of diffusion models, such as the noise schedule or the reverse sampler. In this work, we systematically assess the generalization performance of a diffusion-based speech enhancement model by using multiple speech, noise and binaural room impulse response (BRIR) databases to simulate mismatched acoustic conditions. We also experiment with a noise schedule and a sampler that have not been applied to speech enhancement before. We show that the proposed system substantially benefits from using multiple databases for training, and achieves superior performance compared to state-of-the-art discriminative models in both matched and mismatched conditions. We also show that a Heun-based sampler achieves superior performance at a smaller computational cost compared to a sampler commonly used for speech enhancement.

Jacob Mørk, Holger Severin Bovbjerg, Gergely Kiss et al.

Voice assistants are now widely available, and to activate them a keyword spotting (KWS) algorithm is used. Modern KWS systems are mainly trained using supervised learning methods and require a large amount of labelled data to achieve a good performance. Leveraging unlabelled data through self-supervised learning (SSL) has been shown to increase the accuracy in clean conditions. This paper explores how SSL pretraining such as Data2Vec can be used to enhance the robustness of KWS models in noisy conditions, which is under-explored. Models of three different sizes are pretrained using different pretraining approaches and then fine-tuned for KWS. These models are then tested and compared to models trained using two baseline supervised learning methods, one being standard training using clean data and the other one being multi-style training (MTR). The results show that pretraining and fine-tuning on clean data is superior to supervised learning on clean data across all testing conditions, and superior to supervised MTR for testing conditions of SNR above 5 dB. This indicates that pretraining alone can increase the model's robustness. Finally, it is found that using noisy data for pretraining models, especially with the Data2Vec-denoising approach, significantly enhances the robustness of KWS models in noisy conditions.

Nikolai Lund Kühne, Jan Østergaard, Jesper Jensen et al.

While attention-based architectures, such as Conformers, excel in speech enhancement, they face challenges such as scalability with respect to input sequence length. In contrast, the recently proposed Extended Long Short-Term Memory (xLSTM) architecture offers linear scalability. However, xLSTM-based models remain unexplored for speech enhancement. This paper introduces xLSTM-SENet, the first xLSTM-based single-channel speech enhancement system. A comparative analysis reveals that xLSTM-and notably, even LSTM-can match or outperform state-of-the-art Mamba- and Conformer-based systems across various model sizes in speech enhancement on the VoiceBank+Demand dataset. Through ablation studies, we identify key architectural design choices such as exponential gating and bidirectionality contributing to its effectiveness. Our best xLSTM-based model, xLSTM-SENet2, outperforms state-of-the-art Mamba- and Conformer-based systems of similar complexity on the Voicebank+DEMAND dataset.

Holger Severin Bovbjerg, Jesper Jensen, Jan Østergaard et al.

In this paper, we propose the use of self-supervised pretraining on a large unlabelled data set to improve the performance of a personalized voice activity detection (VAD) model in adverse conditions. We pretrain a long short-term memory (LSTM)-encoder using the autoregressive predictive coding (APC) framework and fine-tune it for personalized VAD. We also propose a denoising variant of APC, with the goal of improving the robustness of personalized VAD. The trained models are systematically evaluated on both clean speech and speech contaminated by various types of noise at different SNR-levels and compared to a purely supervised model. Our experiments show that self-supervised pretraining not only improves performance in clean conditions, but also yields models which are more robust to adverse conditions compared to purely supervised learning.

Deividas Eringis, John Leth, Zheng-Hua Tan et al.

In this paper, we derive a PAC-Bayes bound on the generalisation gap, in a supervised time-series setting for a special class of discrete-time non-linear dynamical systems. This class includes stable recurrent neural networks (RNN), and the motivation for this work was its application to RNNs. In order to achieve the results, we impose some stability constraints, on the allowed models. Here, stability is understood in the sense of dynamical systems. For RNNs, these stability conditions can be expressed in terms of conditions on the weights. We assume the processes involved are essentially bounded and the loss functions are Lipschitz. The proposed bound on the generalisation gap depends on the mixing coefficient of the data distribution, and the essential supremum of the data. Furthermore, the bound converges to zero as the dataset size increases. In this paper, we 1) formalize the learning problem, 2) derive a PAC-Bayesian error bound for such systems, 3) discuss various consequences of this error bound, and 4) show an illustrative example, with discussions on computing the proposed bound. Unlike other available bounds the derived bound holds for non i.i.d. data (time-series) and it does not grow with the number of steps of the RNN.

Holger Severin Bovbjerg, Jan Østergaard, Jesper Jensen et al.

Recently, deep representation learning has shown strong performance in multiple audio tasks. However, its use for learning spatial representations from multichannel audio is underexplored. We investigate the use of a pretraining stage based on feature distillation to learn a robust spatial representation of binaural speech without the need for data labels. In this framework, spatial features are computed from clean binaural speech samples to form prediction labels. These clean features are then predicted from corresponding augmented speech using a neural network. After pretraining, we throw away the spatial feature predictor and use the learned encoder weights to initialize a DoA estimation model which we fine-tune for DoA estimation. Our experiments demonstrate that the pretrained models show improved performance in noisy and reverberant environments after fine-tuning for direction-of-arrival estimation, when compared to fully supervised models and classic signal processing methods.

Yuying Xie, Zheng-Hua Tan

Recent advancements in deep learning have significantly impacted the field of speech signal processing, particularly in the analysis and manipulation of complex spectrograms. This survey provides a comprehensive overview of the state-of-the-art techniques leveraging deep neural networks for processing complex spectrograms, which encapsulate both magnitude and phase information. We begin by introducing complex spectrograms and their associated features for various speech processing tasks. Next, we examine the key components and architectures of complex-valued neural networks, which are specifically designed to handle complex-valued data and have been applied to complex spectrogram processing. As recent studies have primarily focused on applying real-valued neural networks to complex spectrograms, we revisit these approaches and their architectural designs. We then discuss various training strategies and loss functions tailored for training neural networks to process and model complex spectrograms. The survey further examines key applications, including phase retrieval, speech enhancement, and speaker separation, where deep learning has achieved significant progress by leveraging complex spectrograms or their derived feature representations. Additionally, we examine the intersection of complex spectrograms with generative models. This survey aims to serve as a valuable resource for researchers and practitioners in the field of speech signal processing, deep learning and related fields.

Nikolai Lund Kühne, Jesper Jensen, Jan Østergaard et al.

With the advent of new sequence models like Mamba and xLSTM, several studies have shown that these models match or outperform state-of-the-art models in single-channel speech enhancement, automatic speech recognition, and self-supervised audio representation learning. However, prior research has demonstrated that sequence models like LSTM and Mamba tend to overfit to the training set. To address this issue, previous works have shown that adding self-attention to LSTMs substantially improves generalization performance for single-channel speech enhancement. Nevertheless, neither the concept of hybrid Mamba and time-frequency attention models nor their generalization performance have been explored for speech enhancement. In this paper, we propose a novel hybrid architecture, MambAttention, which combines Mamba and shared time- and frequency-multi-head attention modules for generalizable single-channel speech enhancement. To train our model, we introduce VoiceBank+Demand Extended (VB-DemandEx), a dataset inspired by VoiceBank+Demand but with more challenging noise types and lower signal-to-noise ratios. Trained on VB-DemandEx, our proposed MambAttention model significantly outperforms existing state-of-the-art LSTM-, xLSTM-, Mamba-, and Conformer-based systems of similar complexity across all reported metrics on two out-of-domain datasets: DNS 2020 and EARS-WHAM_v2, while matching their performance on the in-domain dataset VB-DemandEx. Ablation studies highlight the role of weight sharing between the time- and frequency-multi-head attention modules for generalization performance. Finally, we explore integrating the shared time- and frequency-multi-head attention modules with LSTM and xLSTM, which yields a notable performance improvement on the out-of-domain datasets. However, our MambAttention model remains superior on both out-of-domain datasets across all reported evaluation metrics.

Holger Severin Bovbjerg, Jan Østergaard, Jesper Jensen et al.

Target-Speaker Voice Activity Detection (TS-VAD) is the task of detecting the presence of speech from a known target-speaker in an audio frame. Recently, deep neural network-based models have shown good performance in this task. However, training these models requires extensive labelled data, which is costly and time-consuming to obtain, particularly if generalization to unseen environments is crucial. To mitigate this, we propose a causal, Self-Supervised Learning (SSL) pretraining framework, called Denoising Autoregressive Predictive Coding (DN-APC), to enhance TS-VAD performance in noisy conditions. We also explore various speaker conditioning methods and evaluate their performance under different noisy conditions. Our experiments show that DN-APC improves performance in noisy conditions, with a general improvement of approx. 2% in both seen and unseen noise. Additionally, we find that FiLM conditioning provides the best overall performance. Representation analysis via tSNE plots reveals robust initial representations of speech and non-speech from pretraining. This underscores the effectiveness of SSL pretraining in improving the robustness and performance of TS-VAD models in noisy environments.

Nikolai Lund Kühne, Jesper Jensen, Jan Østergaard et al.

Recent advances in speech enhancement have shown that models combining Mamba and attention mechanisms yield superior cross-corpus generalization performance. At the same time, integrating Mamba in a U-Net structure has yielded state-of-the-art enhancement performance, while reducing both model size and computational complexity. Inspired by these insights, we propose RWSA-MambaUNet, a novel and efficient hybrid model combining Mamba and multi-head attention in a U-Net structure for improved cross-corpus performance. Resolution-wise shared attention (RWSA) refers to layerwise attention-sharing across corresponding time- and frequency resolutions. Our best-performing RWSA-MambaUNet model achieves state-of-the-art generalization performance on two out-of-domain test sets. Notably, our smallest model surpasses all baselines on the out-of-domain DNS 2020 test set in terms of PESQ, SSNR, and ESTOI, and on the out-of-domain EARS-WHAM_v2 test set in terms of SSNR, ESTOI, and SI-SDR, while using less than half the model parameters and a fraction of the FLOPs.

Lasse Borgholt, Jakob Havtorn, Christian Igel et al.

Modern neural networks have greatly improved performance across speech recognition benchmarks. However, gains are often driven by frequent words with limited semantic weight, which can obscure meaningful differences in word error rate, the primary evaluation metric. Errors in rare terms, named entities, and domain-specific vocabulary are more consequential, but remain hidden by aggregate metrics. This highlights the need for finer-grained error analysis, which depends on accurate alignment between reference and model transcripts. However, conventional alignment methods are not designed for such precision. We propose a novel alignment algorithm that couples dynamic programming with beam search scoring. Compared to traditional text alignment methods, our approach provides more accurate alignment of individual errors, enabling reliable error analysis. The algorithm is made available via PyPI.

Sarthak Yadav, Sergios Theodoridis, Zheng-Hua Tan

In recent years, self-supervised learning has amassed significant interest for training deep neural representations without labeled data. One such self-supervised learning approach is masked spectrogram modeling, where the objective is to learn semantically rich contextual representations by predicting removed or hidden portions of the input audio spectrogram. With the Transformer neural architecture at its core, masked spectrogram modeling has emerged as the prominent approach for learning general purpose audio representations, a.k.a. audio foundation models. Meanwhile, addressing the issues of the Transformer architecture, in particular the underlying Scaled Dot-product Attention operation, which scales quadratically with input sequence length, has led to renewed interest in recurrent sequence modeling approaches. Among them, Selective structured state space models (such as Mamba) and extended Long Short-Term Memory (xLSTM) are the two most promising approaches which have experienced widespread adoption. While the body of work on these two topics continues to grow, there is currently a lack of an adequate overview encompassing the intersection of these topics. In this paper, we present a comprehensive overview of the aforementioned research domains, covering masked spectrogram modeling and the previously mentioned neural sequence modeling architectures, Mamba and xLSTM. Further, we compare Transformers, Mamba and xLSTM based masked spectrogram models in a unified, reproducible framework on ten diverse downstream audio classification tasks, which will help interested readers to make informed decisions regarding suitability of the evaluated approaches to adjacent applications.

Kevin Wilkinghoff, Zheng-Hua Tan

Reasoning about spatial audio with large language models requires a spatial audio encoder as an acoustic front-end to obtain audio embeddings for further processing. Such an encoder needs to capture all information required to detect the type of sound events, as well as the direction and distance of their corresponding sources. Accomplishing this with a single audio encoder is demanding as the information required for each of these tasks is mostly independent of each other. As a result, the performance obtained with a single encoder is often worse than when using task-specific audio encoders. In this work, we present DSpAST, a novel audio encoder based on SpatialAST that learns disentangled representations of spatial audio while having only 0.2% additional parameters. Experiments on SpatialSoundQA with the spatial audio reasoning system BAT demonstrate that DSpAST significantly outperforms SpatialAST.

Sarthak Yadav, Sergios Theodoridis, Zheng-Hua Tan

Masked Autoencoders (MAEs) trained on audio spectrogram patches have emerged as a prominent approach for learning self-supervised audio representations. While several recent papers have evaluated key aspects of training MAEs on audio data, the majority of these approaches still leverage vanilla transformer building blocks, whereas the transformer community has seen steady integration of newer architectural advancements. In this work, we propose AudioMAE++, a revamped audio masked autoencoder with two such enhancements, namely macaron-style transformer blocks with gated linear units. When pretrained on the AudioSet dataset, the proposed AudioMAE++ models outperform existing MAE based approaches on 10 diverse downstream tasks, demonstrating excellent performance on audio classification and speech-based benchmarks. The proposed AudioMAE++ models also demonstrate excellent scaling characteristics, outperforming directly comparable standard MAE baselines with up to 4x more parameters.

Achintya kr. Sarkar, Priyanka Dwivedi, Zheng-Hua Tan

In this paper, we propose several methods that incorporate vocal tract length (VTL) warped features for spoken keyword spotting (KWS). The first method, VTL-independent KWS, involves training a single deep neural network (DNN) that utilizes VTL features with various warping factors. During training, a specific VTL feature is randomly selected per epoch, allowing the exploration of VTL variations. During testing, the VTL features with different warping factors of a test utterance are scored against the DNN and combined with equal weight. In the second method scores the conventional features of a test utterance (without VTL warping) against the DNN. The third method, VTL-concatenation KWS, concatenates VTL warped features to form high-dimensional features for KWS. Evaluations carried out on the English Google Command dataset demonstrate that the proposed methods improve the accuracy of KWS.

Sarthak Yadav, Zheng-Hua Tan

Despite its widespread adoption as the prominent neural architecture, the Transformer has spurred several independent lines of work to address its limitations. One such approach is selective state space models, which have demonstrated promising results for language modelling. However, their feasibility for learning self-supervised, general-purpose audio representations is yet to be investigated. This work proposes Audio Mamba, a selective state space model for learning general-purpose audio representations from randomly masked spectrogram patches through self-supervision. Empirical results on ten diverse audio recognition downstream tasks show that the proposed models, pretrained on the AudioSet dataset, consistently outperform comparable self-supervised audio spectrogram transformer (SSAST) baselines by a considerable margin and demonstrate better performance in dataset size, sequence length and model size comparisons.

Fan Yu, Shiliang Zhang, Pengcheng Guo et al.

The ICASSP 2022 Multi-channel Multi-party Meeting Transcription Grand Challenge (M2MeT) focuses on one of the most valuable and the most challenging scenarios of speech technologies. The M2MeT challenge has particularly set up two tracks, speaker diarization (track 1) and multi-speaker automatic speech recognition (ASR) (track 2). Along with the challenge, we released 120 hours of real-recorded Mandarin meeting speech data with manual annotation, including far-field data collected by 8-channel microphone array as well as near-field data collected by each participants' headset microphone. We briefly describe the released dataset, track setups, baselines and summarize the challenge results and major techniques used in the submissions.

Achintya kr. Sarkar, Zheng-Hua Tan

Deep representation learning has gained significant momentum in advancing text-dependent speaker verification (TD-SV) systems. When designing deep neural networks (DNN) for extracting bottleneck features, key considerations include training targets, activation functions, and loss functions. In this paper, we systematically study the impact of these choices on the performance of TD-SV. For training targets, we consider speaker identity, time-contrastive learning (TCL) and auto-regressive prediction coding with the first being supervised and the last two being self-supervised. Furthermore, we study a range of loss functions when speaker identity is used as the training target. With regard to activation functions, we study the widely used sigmoid function, rectified linear unit (ReLU), and Gaussian error linear unit (GELU). We experimentally show that GELU is able to reduce the error rates of TD-SV significantly compared to sigmoid, irrespective of training target. Among the three training targets, TCL performs the best. Among the various loss functions, cross entropy, joint-softmax and focal loss functions outperform the others. Finally, score-level fusion of different systems is also able to reduce the error rates. Experiments are conducted on the RedDots 2016 challenge database for TD-SV using short utterances. For the speaker classifications, the well-known Gaussian mixture model-universal background model (GMM-UBM) and i-vector techniques are used.

Iván López-Espejo, Zheng-Hua Tan, John Hansen et al.

Spoken keyword spotting (KWS) deals with the identification of keywords in audio streams and has become a fast-growing technology thanks to the paradigm shift introduced by deep learning a few years ago. This has allowed the rapid embedding of deep KWS in a myriad of small electronic devices with different purposes like the activation of voice assistants. Prospects suggest a sustained growth in terms of social use of this technology. Thus, it is not surprising that deep KWS has become a hot research topic among speech scientists, who constantly look for KWS performance improvement and computational complexity reduction. This context motivates this paper, in which we conduct a literature review into deep spoken KWS to assist practitioners and researchers who are interested in this technology. Specifically, this overview has a comprehensive nature by covering a thorough analysis of deep KWS systems (which includes speech features, acoustic modeling and posterior handling), robustness methods, applications, datasets, evaluation metrics, performance of deep KWS systems and audio-visual KWS. The analysis performed in this paper allows us to identify a number of directions for future research, including directions adopted from automatic speech recognition research and directions that are unique to the problem of spoken KWS.

Andreas Jonas Fuglsig, Jan Østergaard, Jesper Jensen et al.

This paper considers speech enhancement of signals picked up in one noisy environment which must be presented to a listener in another noisy environment. Recently, it has been shown that an optimal solution to this problem requires the consideration of the noise sources in both environments jointly. However, the existing optimal mutual information based method requires a complicated system model that includes natural speech variations, and relies on approximations and assumptions of the underlying signal distributions. In this paper, we propose to use a simpler signal model and optimize speech intelligibility based on the Approximated Speech Intelligibility Index (ASII). We derive a closed-form solution to the joint far- and near-end speech enhancement problem that is independent of the marginal distribution of signal coefficients, and that achieves similar performance to existing work. In addition, we do not need to model or optimize for natural speech variations.

Deividas Eringis, John Leth, Zheng-Hua Tan et al.

In this short article, we showcase the derivation of the optimal (minimum error variance) estimator, when one part of the stochastic LTI system output is not measured but is able to be predicted from the measured system outputs. Similar derivations have been done before but not using state-space representation.

Max Væhrens, Andreas Jonas Fuglsig, Anders Post Jacobsen et al.

Voice activity detection (VAD) remains a challenge in noisy environments. With access to multiple microphones, prior studies have attempted to improve the noise robustness of VAD by creating multi-channel VAD (MVAD) methods. However, MVAD is relatively new compared to single-channel VAD (SVAD), which has been thoroughly developed in the past. It might therefore be advantageous to improve SVAD methods with pre-processing to obtain superior VAD, which is under-explored. This paper improves SVAD through two pre-processing methods, a beamformer and a spatial target speaker detector. The spatial detector sets signal frames to zero when no potential speaker is present within a target direction. The detector may be implemented as a filter, meaning the input signal for the SVAD is filtered according to the detector's output; or it may be implemented as a spatial VAD to be combined with the SVAD output. The evaluation is made on a noisy reverberant speech database, with clean speech from the Aurora 2 database and with white and babble noise. The results show that SVAD algorithms are significantly improved by the presented pre-processing methods, especially the spatial detector, across all signal-to-noise ratios. The SVAD algorithms with pre-processing significantly outperform a baseline MVAD in challenging noise conditions.

Morten Kolbæk, Zheng-Hua Tan, Søren Holdt Jensen et al.

In recent years, speech processing algorithms have seen tremendous progress primarily due to the deep learning renaissance. This is especially true for speech separation where the time-domain audio separation network (TasNet) has led to significant improvements. However, for the related task of single-speaker speech enhancement, which is of obvious importance, it is yet unknown, if the TasNet architecture is equally successful. In this paper, we show that TasNet improves state-of-the-art also for speech enhancement, and that the largest gains are achieved for modulated noise sources such as speech. Furthermore, we show that TasNet learns an efficient inner-domain representation, where target and noise signal components are highly separable. This is especially true for noise in terms of interfering speech signals, which might explain why TasNet performs so well on the separation task. Additionally, we show that TasNet performs poorly for large frame hops and conjecture that aliasing might be the main cause of this performance drop. Finally, we show that TasNet consistently outperforms a state-of-the-art single-speaker speech enhancement system.

Deividas Eringis, John Leth, Zheng-Hua Tan et al.

In this paper we derive a PAC-Bayesian error bound for autonomous stochastic LTI state-space models. The motivation for deriving such error bounds is that they will allow deriving similar error bounds for more general dynamical systems, including recurrent neural networks. In turn, PACBayesian error bounds are known to be useful for analyzing machine learning algorithms and for deriving new ones.

Achintya Kumar Sarkar, Md Sahidullah, Zheng-Hua Tan

In this paper, we propose a novel method that trains pass-phrase specific deep neural network (PP-DNN) based auto-encoders for creating augmented data for text-dependent speaker verification (TD-SV). Each PP-DNN auto-encoder is trained using the utterances of a particular pass-phrase available in the target enrollment set with two methods: (i) transfer learning and (ii) training from scratch. Next, feature vectors of a given utterance are fed to the PP-DNNs and the output from each PP-DNN at frame-level is considered one new set of generated data. The generated data from each PP-DNN is then used for building a TD-SV system in contrast to the conventional method that considers only the evaluation data available. The proposed approach can be considered as the transformation of data to the pass-phrase specific space using a non-linear transformation learned by each PP-DNN. The method develops several TD-SV systems with the number equal to the number of PP-DNNs separately trained for each pass-phrases for the evaluation. Finally, the scores of the different TD-SV systems are fused for decision making. Experiments are conducted on the RedDots challenge 2016 database for TD-SV using short utterances. Results show that the proposed method improves the performance for both conventional cepstral feature and deep bottleneck feature using both Gaussian mixture model - universal background model (GMM-UBM) and i-vector framework.

Achintya kr. Sarkar, Zheng-Hua Tan

In this letter, we propose a vocal tract length (VTL) perturbation method for text-dependent speaker verification (TD-SV), in which a set of TD-SV systems are trained, one for each VTL factor, and score-level fusion is applied to make a final decision. Next, we explore the bottleneck (BN) feature extracted by training deep neural networks with a self-supervised objective, autoregressive predictive coding (APC), for TD-SV and compare it with the well-studied speaker-discriminant BN feature. The proposed VTL method is then applied to APC and speaker-discriminant BN features. In the end, we combine the VTL perturbation systems trained on MFCC and the two BN features in the score domain. Experiments are performed on the RedDots challenge 2016 database of TD-SV using short utterances with Gaussian mixture model-universal background model and i-vector techniques. Results show the proposed methods significantly outperform the baselines.

Cristian J. Vaca-Rubio, Pablo Ramirez-Espinosa, Kimmo Kansanen et al.

Sensing capability is one of the most highlighted new feature of future 6G wireless networks. This paper addresses the sensing potential of Large Intelligent Surfaces (LIS) in an exemplary Industry 4.0 scenario. Besides the attention received by LIS in terms of communication aspects, it can offer a high-resolution rendering of the propagation environment. This is because, in an indoor setting, it can be placed in proximity to the sensed phenomena, while the high resolution is offered by densely spaced tiny antennas deployed over a large area. By treating an LIS as a radio image of the environment relying on the received signal power, we develop techniques to sense the environment, by leveraging the tools of image processing and machine learning. Once a holographic image is obtained, a Denoising Autoencoder (DAE) network can be used for constructing a super-resolution image leading to sensing advantages not available in traditional sensing systems. Also, we derive a statistical test based on the Generalized Likelihood Ratio (GLRT) as a benchmark for the machine learning solution. We test these methods for a scenario where we need to detect whether an industrial robot deviates from a predefined route. The results show that the LIS-based sensing offers high precision and has a high application potential in indoor industrial environments.

Zeyu Song, Dongliang Chang, Zhanyu Ma et al.

The loss function is a key component in deep learning models. A commonly used loss function for classification is the cross entropy loss, which is a simple yet effective application of information theory for classification problems. Based on this loss, many other loss functions have been proposed,~\emph{e.g.}, by adding intra-class and inter-class constraints to enhance the discriminative ability of the learned features. However, these loss functions fail to consider the connections between the feature distribution and the model structure. Aiming at addressing this problem, we propose a channel correlation loss (CC-Loss) that is able to constrain the specific relations between classes and channels as well as maintain the intra-class and the inter-class separability. CC-Loss uses a channel attention module to generate channel attention of features for each sample in the training stage. Next, an Euclidean distance matrix is calculated to make the channel attention vectors associated with the same class become identical and to increase the difference between different classes. Finally, we obtain a feature embedding with good intra-class compactness and inter-class separability.Experimental results show that two different backbone models trained with the proposed CC-Loss outperform the state-of-the-art loss functions on three image classification datasets.

Giovanni Morrone, Daniel Michelsanti, Zheng-Hua Tan et al.

In this paper, we present a deep-learning-based framework for audio-visual speech inpainting, i.e., the task of restoring the missing parts of an acoustic speech signal from reliable audio context and uncorrupted visual information. Recent work focuses solely on audio-only methods and generally aims at inpainting music signals, which show highly different structure than speech. Instead, we inpaint speech signals with gaps ranging from 100 ms to 1600 ms to investigate the contribution that vision can provide for gaps of different duration. We also experiment with a multi-task learning approach where a phone recognition task is learned together with speech inpainting. Results show that the performance of audio-only speech inpainting approaches degrades rapidly when gaps get large, while the proposed audio-visual approach is able to plausibly restore missing information. In addition, we show that multi-task learning is effective, although the largest contribution to performance comes from vision.

Daniel Michelsanti, Zheng-Hua Tan, Shi-Xiong Zhang et al.

Speech enhancement and speech separation are two related tasks, whose purpose is to extract either one or more target speech signals, respectively, from a mixture of sounds generated by several sources. Traditionally, these tasks have been tackled using signal processing and machine learning techniques applied to the available acoustic signals. Since the visual aspect of speech is essentially unaffected by the acoustic environment, visual information from the target speakers, such as lip movements and facial expressions, has also been used for speech enhancement and speech separation systems. In order to efficiently fuse acoustic and visual information, researchers have exploited the flexibility of data-driven approaches, specifically deep learning, achieving strong performance. The ceaseless proposal of a large number of techniques to extract features and fuse multimodal information has highlighted the need for an overview that comprehensively describes and discusses audio-visual speech enhancement and separation based on deep learning. In this paper, we provide a systematic survey of this research topic, focusing on the main elements that characterise the systems in the literature: acoustic features; visual features; deep learning methods; fusion techniques; training targets and objective functions. In addition, we review deep-learning-based methods for speech reconstruction from silent videos and audio-visual sound source separation for non-speech signals, since these methods can be more or less directly applied to audio-visual speech enhancement and separation. Finally, we survey commonly employed audio-visual speech datasets, given their central role in the development of data-driven approaches, and evaluation methods, because they are generally used to compare different systems and determine their performance.

Md Sahidullah, Achintya Kumar Sarkar, Ville Vestman et al.

In this work, we present the system description of the UIAI entry for the short-duration speaker verification (SdSV) challenge 2020. Our focus is on Task 1 dedicated to text-dependent speaker verification. We investigate different feature extraction and modeling approaches for automatic speaker verification (ASV) and utterance verification (UV). We have also studied different fusion strategies for combining UV and ASV modules. Our primary submission to the challenge is the fusion of seven subsystems which yields a normalized minimum detection cost function (minDCF) of 0.072 and an equal error rate (EER) of 2.14% on the evaluation set. The single system consisting of a pass-phrase identification based model with phone-discriminative bottleneck features gives a normalized minDCF of 0.118 and achieves 19% relative improvement over the state-of-the-art challenge baseline.

Achintya Kumar Sarkar, Himangshu Sarma, Priyanka Dwivedi et al.

Data augmentation is commonly used for generating additional data from the available training data to achieve a robust estimation of the parameters of complex models like the one for speaker verification (SV), especially for under-resourced applications. SV involves training speaker-independent (SI) models and speaker-dependent models where speakers are represented by models derived from an SI model using the training data for the particular speaker during the enrollment phase. While data augmentation for training SI models is well studied, data augmentation for speaker enrollment is rarely explored. In this paper, we propose the use of data augmentation methods for generating extra data to empower speaker enrollment. Each data augmentation method generates a new data set. Two strategies of using the data sets are explored: the first one is to training separate systems and fuses them at the score level and the other is to conduct multi-conditional training. Furthermore, we study the effect of data augmentation under noisy conditions. Experiments are performed on RedDots challenge 2016 database, and the results validate the effectiveness of the proposed methods.