Emanuël A. P. Habets

Shrishti Saha Shetu, Emanuël A. P. Habets, Andreas Brendel

In this study, we conduct a comprehensive comparative analysis of generative and discriminative deep learning-based speech enhancement methods, specifically in noise reduction tasks. Our investigation focuses on evaluating their effectiveness under high and low signal-to-noise ratio conditions, considering both matched and mismatched training scenarios. We further investigate the impact of training data volume, model convergence speed, and interpret the performance differences in terms of objective results for the considered training paradigms. Additionally, we compare the complexity-performance trade-off and the practical viability of these approaches. To further strengthen the evaluation, we study the hallucination characteristics of generative approaches in terms of word error rate and phoneme similarity. The insights derived from this study provide empirical evidence to assist researchers and practitioners in understanding whether the perceptual gains of different approaches justify their computational cost in practical applications.

Julian Wechsler, Srikanth Raj Chetupalli, Wolfgang Mack et al.

We consider the task of region-based source separation of reverberant multi-microphone recordings. We assume pre-defined spatial regions with a single active source per region. The objective is to estimate the signals from the individual spatial regions as captured by a reference microphone while retaining a correspondence between signals and spatial regions. We propose a data-driven approach using a modified version of a state-of-the-art network, where different layers model spatial and spectro-temporal information. The network is trained to enforce a fixed mapping of regions to network outputs. Using speech from LibriMix, we construct a data set specifically designed to contain the region information. Additionally, we train the network with permutation invariant training. We show that both training methods result in a fixed mapping of regions to network outputs, achieve comparable performance, and that the networks exploit spatial information. The proposed network outperforms a baseline network by 1.5 dB in scale-invariant signal-to-distortion ratio.

Ahmad Aloradi, Tim Roith, Emanuël A. P. Habets et al.

Sparse training reduces the memory and computational costs of deep neural networks. However, sparse optimization methods, e.g., those adding an $\ell_1$ penalty, often control sparsity only indirectly through a regularization parameter $λ$, whose mapping to the final sparsity rate is non-trivial. In our experiments, we found this parameter sensitivity to be particularly pronounced for Bregman-based optimizers. Specifically, the two variants LinBreg and AdaBreg reach the same sparsity at $λ$ values that differ by up to two orders of magnitude, requiring expensive trial-and-error sweeps to achieve a user-specified sparsity. To address this, we propose an adaptive regularization scheme that updates $λ$ based on the difference between the model's current sparsity and the target sparsity. We analyze the resulting algorithm and evaluate it on automatic speaker verification with ECAPA-TDNN and ResNet34 on VoxCeleb and CNCeleb. The proposed method reliably achieves sparsity targets ranging between 75% and 99%. It also converges faster than the oracle-tuned non-adaptive baseline during early training and matches or surpasses its final performance in equal error rate. We further show that the adaptive scheme inherits key properties from its non-adaptive counterpart, including improved out-of-distribution robustness over the dense baselines.

Ünal Ege Gaznepoglu, Anna Leschanowsky, Ahmad Aloradi et al.

Speaker anonymization systems hide the identity of speakers while preserving other information such as linguistic content and emotions. To evaluate their privacy benefits, attacks in the form of automatic speaker verification (ASV) systems are employed. In this study, we assess the impact of intra-speaker linguistic content similarity in the attacker training and evaluation datasets, by adapting BERT, a language model, as an ASV system. On the VoicePrivacy Attacker Challenge datasets, our method achieves a mean equal error rate (EER) of 35%, with certain speakers attaining EERs as low as 2%, based solely on the textual content of their utterances. Our explainability study reveals that the system decisions are linked to semantically similar keywords within utterances, stemming from how LibriSpeech is curated. Our study suggests reworking the VoicePrivacy datasets to ensure a fair and unbiased evaluation and challenge the reliance on global EER for privacy evaluations.

Anna Leschanowsky, Zahra Kolagar, Erion Çano et al.

The transparency principle of the General Data Protection Regulation (GDPR) requires data processing information to be clear, precise, and accessible. While language models show promise in this context, their probabilistic nature complicates truthfulness and comprehensibility. This paper examines state-of-the-art Retrieval Augmented Generation (RAG) systems enhanced with alignment techniques to fulfill GDPR obligations. We evaluate RAG systems incorporating an alignment module like Rewindable Auto-regressive Inference (RAIN) and our proposed multidimensional extension, MultiRAIN, using a Privacy Q&A dataset. Responses are optimized for preciseness and comprehensibility and are assessed through 21 metrics, including deterministic and large language model-based evaluations. Our results show that RAG systems with an alignment module outperform baseline RAG systems on most metrics, though none fully match human answers. Principal component analysis of the results reveals complex interactions between metrics, highlighting the need to refine metrics. This study provides a foundation for integrating advanced natural language processing systems into legal compliance frameworks.

Florian Lux, Sarina Meyer, Lyonel Behringer et al.

In this work, we take on the challenging task of building a single text-to-speech synthesis system that is capable of generating speech in over 7000 languages, many of which lack sufficient data for traditional TTS development. By leveraging a novel integration of massively multilingual pretraining and meta learning to approximate language representations, our approach enables zero-shot speech synthesis in languages without any available data. We validate our system's performance through objective measures and human evaluation across a diverse linguistic landscape. By releasing our code and models publicly, we aim to empower communities with limited linguistic resources and foster further innovation in the field of speech technology.

Philipp Götz, Cagdas Tuna, Andreas Walther et al.

The estimation of reverberation time from real-world signals plays a central role in a wide range of applications. In many scenarios, acoustic conditions change over time which in turn requires the estimate to be updated continuously. Previously proposed methods involving deep neural networks were mostly designed and tested under the assumption of static acoustic conditions. In this work, we show that these approaches can perform poorly in dynamically evolving acoustic environments. Motivated by a recent trend towards data-centric approaches in machine learning, we propose a novel way of generating training data and demonstrate, using an existing deep neural network architecture, the considerable improvement in the ability to follow temporal changes in reverberation time.

Mohamed Elminshawi, Wolfgang Mack, Srikanth Raj Chetupalli et al.

Speaker extraction (SE) aims to segregate the speech of a target speaker from a mixture of interfering speakers with the help of auxiliary information. Several forms of auxiliary information have been employed in single-channel SE, such as a speech snippet enrolled from the target speaker or visual information corresponding to the spoken utterance. The effectiveness of the auxiliary information in SE is typically evaluated by comparing the extraction performance of SE with uninformed speaker separation (SS) methods. Following this evaluation protocol, many SE studies have reported performance improvement compared to SS, attributing this to the auxiliary information. However, such studies have been conducted on a few datasets and have not considered recent deep neural network architectures for SS that have shown impressive separation performance. In this paper, we examine the role of the auxiliary information in SE for different input scenarios and over multiple datasets. Specifically, we compare the performance of two SE systems (audio-based and video-based) with SS using a common framework that utilizes the recently proposed dual-path recurrent neural network as the main learning machine. Experimental evaluation on various datasets demonstrates that the use of auxiliary information in the considered SE systems does not always lead to better extraction performance compared to the uninformed SS system. Furthermore, we offer insights into the behavior of the SE systems when provided with different and distorted auxiliary information given the same mixture input.

Wolfgang Mack, Julian Wechsler, Emanuël A. P. Habets

The direction-of-arrival (DOA) of sound sources is an essential acoustic parameter used, e.g., for multi-channel speech enhancement or source tracking. Complex acoustic scenarios consisting of sources-of-interest, interfering sources, reverberation, and noise make the estimation of the DOAs corresponding to the sources-of-interest a challenging task. Recently proposed attention mechanisms allow DOA estimators to focus on the sources-of-interest and disregard interference and noise, i.e., they are signal-aware. The attention is typically obtained by a deep neural network (DNN) from a short-time Fourier transform (STFT) based representation of a single microphone signal. Subsequently, attention has been applied as binary or ratio weighting to STFT-based microphone signal representations to reduce the impact of frequency bins dominated by noise, interference, or reverberation. The impact of attention on DOA estimators and different training strategies for attention and DOA DNNs are not yet studied in depth. In this paper, we evaluate systems consisting of different DNNs and signal processing-based methods for DOA estimation when attention is applied. Additionally, we propose training strategies for attention-based DOA estimation optimized via a DOA objective, i.e., end-to-end. The evaluation of the proposed and the baseline systems is performed using data generated with simulated and measured room impulse responses under various acoustic conditions, like reverberation times, noise, and source array distances. Overall, DOA estimation using attention in combination with signal-processing methods exhibits a far lower computational complexity than a fully DNN-based system; however, it yields comparable results.

Mohamed Elminshawi, Wolfgang Mack, Emanuël A. P. Habets

Informed speaker extraction aims to extract a target speech signal from a mixture of sources given prior knowledge about the desired speaker. Recent deep learning-based methods leverage a speaker discriminative model that maps a reference snippet uttered by the target speaker into a single embedding vector that encapsulates the characteristics of the target speaker. However, such modeling deliberately neglects the time-varying properties of the reference signal. In this work, we assume that a reference signal is available that is temporally correlated with the target signal. To take this correlation into account, we propose a time-varying source discriminative model that captures the temporal dynamics of the reference signal. We also show that existing methods and the proposed method can be generalized to non-speech sources as well. Experimental results demonstrate that the proposed method significantly improves the extraction performance when applied in an acoustic echo reduction scenario.

Fabian Hübner, Wolfgang Mack, Emanuël A. P. Habets

Deep learning (DL) based direction of arrival (DOA) estimation is an active research topic and currently represents the state-of-the-art. Usually, DL-based DOA estimators are trained with recorded data or computationally expensive generated data. Both data types require significant storage and excessive time to, respectively, record or generate. We propose a low complexity online data generation method to train DL models with a phase-based feature input. The data generation method models the phases of the microphone signals in the frequency domain by employing a deterministic model for the direct path and a statistical model for the late reverberation of the room transfer function. By an evaluation using data from measured room impulse responses, we demonstrate that a model trained with the proposed training data generation method performs comparably to models trained with data generated based on the source-image method.

Shrishti Saha Shetu, Soumitro Chakrabarty, Emanuël A. P. Habets

Audio-visual speech enhancement (AVSE) methods use both audio and visual features for the task of speech enhancement and the use of visual features has been shown to be particularly effective in multi-speaker scenarios. In the majority of deep neural network (DNN) based AVSE methods, the audio and visual data are first processed separately using different sub-networks, and then the learned features are fused to utilize the information from both modalities. There have been various studies on suitable audio input features and network architectures, however, to the best of our knowledge, there is no published study that has investigated which visual features are best suited for this specific task. In this work, we perform an empirical study of the most commonly used visual features for DNN based AVSE, the pre-processing requirements for each of these features, and investigate their influence on the performance. Our study shows that despite the overall better performance of embedding-based features, their computationally intensive pre-processing make their use difficult in low resource systems. For such systems, optical flow or raw pixels-based features might be better suited.

Alessandro Ilic Mezza, Emanuël A. P. Habets, Meinard Müller et al.

The performance of machine learning algorithms is known to be negatively affected by possible mismatches between training (source) and test (target) data distributions. In fact, this problem emerges whenever an acoustic scene classification system which has been trained on data recorded by a given device is applied to samples acquired under different acoustic conditions or captured by mismatched recording devices. To address this issue, we propose an unsupervised domain adaptation method that consists of aligning the first- and second-order sample statistics of each frequency band of target-domain acoustic scenes to the ones of the source-domain training dataset. This model-agnostic approach is devised to adapt audio samples from unseen devices before they are fed to a pre-trained classifier, thus avoiding any further learning phase. Using the DCASE 2018 Task 1-B development dataset, we show that the proposed method outperforms the state-of-the-art unsupervised methods found in the literature in terms of both source- and target-domain classification accuracy.

Sebastian J. Schlecht, Emanuël A. P. Habets

Feedback delay networks (FDNs) are recursive filters, which are widely used for artificial reverberation and decorrelation. One central challenge in the design of FDNs is the generation of sufficient echo density in the impulse response without compromising the computational efficiency. In a previous contribution, we have demonstrated that the echo density of an FDN can be increased by introducing so-called delay feedback matrices where each matrix entry is a scalar gain and a delay. In this contribution, we generalize the feedback matrix to arbitrary lossless filter feedback matrices (FFMs). As a special case, we propose the velvet feedback matrix, which can create dense impulse responses at a minimal computational cost. Further, FFMs can be used to emulate the scattering effects of non-specular reflections. We demonstrate the effectiveness of FFMs in terms of echo density and modal distribution.

Wolfgang Mack, Emanuël A. P. Habets

Signal extraction from a single-channel mixture with additional undesired signals is most commonly performed using time-frequency (TF) masks. Typically, the mask is estimated with a deep neural network (DNN), and element-wise applied to the complex mixture short-time Fourier transform (STFT) representation to perform the extraction. Ideal mask magnitudes are zero for solely undesired signals in a TF bin and undefined for total destructive interference. Usually, masks have an upper bound to provide well-defined DNN outputs at the cost of limited extraction capabilities. We propose to estimate with a DNN a complex TF filter for each mixture TF bin which maps an STFT area in the respective mixture to the desired TF bin to address destructive interference in mixture TF bins. The DNN is optimized by minimizing the error between the extracted and the ground-truth desired signal allowing to learn the TF filters without having to specify ground-truth TF filters. We compare our approach with complex and real-valued TF masks by separating speech from a variety of different sound and noise classes from the Google AudioSet corpus. We also process the mixture STFT with notch-filters and zero whole time-frames, to simulate packet-loss during transmission, to demonstrate the reconstruction capabilities of our approach. The proposed method outperformed the baselines, especially when notch-filters and time-frame zeroing were applied.

Soumitro Chakrabarty, Emanuël A. P. Habets

In a recent work on direction-of-arrival (DOA) estimation of multiple speakers with convolutional neural networks (CNNs), the phase component of short-time Fourier transform (STFT) coefficients of the microphone signal is given as input and small filters are used to learn the phase relations between neighboring microphones. Due to this chosen filter size, $M-1$ convolution layers are required to achieve the best performance for a microphone array with M microphones. For arrays with large number of microphones, this requirement leads to a high computational cost making the method practically infeasible. In this work, we propose to use systematic dilations of the convolution filters in each of the convolution layers of the previously proposed CNN for expansion of the receptive field of the filters to reduce the computational cost of the method. Different strategies for expansion of the receptive field of the filters for a specific microphone array are explored. With experimental analysis of the different strategies, it is shown that an aggressive expansion strategy results in a considerable reduction in computational cost while a relatively gradual expansion of the receptive field exhibits the best DOA estimation performance along with reduction in the computational cost.

Daniele Mirabilii, Emanuël A. P. Habets

The difference-to-sum power ratio was proposed and used to suppress wind noise under specific acoustic conditions. In this contribution, a general formulation of the difference-to-sum power ratio associated with a mixture of speech and wind noise is proposed and analyzed. In particular, it is assumed that the complex coherence of convective turbulence can be modelled by the Corcos model. In contrast to the work in which the power ratio was first presented, the employed Corcos model holds for every possible air stream direction and takes into account the lateral coherence decay rate. The obtained expression is subsequently validated with real data for a dual microphone set-up. Finally, the difference-to- sum power ratio is exploited as a spatial feature to indicate the frame-wise presence of wind noise, obtaining improved detection performance when compared to an existing multi-channel wind noise detection approach.

Soumitro Chakrabarty, Emanuël A. P. Habets

Supervised learning based methods for source localization, being data driven, can be adapted to different acoustic conditions via training and have been shown to be robust to adverse acoustic environments. In this paper, a convolutional neural network (CNN) based supervised learning method for estimating the direction-of-arrival (DOA) of multiple speakers is proposed. Multi-speaker DOA estimation is formulated as a multi-class multi-label classification problem, where the assignment of each DOA label to the input feature is treated as a separate binary classification problem. The phase component of the short-time Fourier transform (STFT) coefficients of the received microphone signals are directly fed into the CNN, and the features for DOA estimation are learnt during training. Utilizing the assumption of disjoint speaker activity in the STFT domain, a novel method is proposed to train the CNN with synthesized noise signals. Through experimental evaluation with both simulated and measured acoustic impulse responses, the ability of the proposed DOA estimation approach to adapt to unseen acoustic conditions and its robustness to unseen noise type is demonstrated. Through additional empirical investigation, it is also shown that with an array of M microphones our proposed framework yields the best localization performance with M-1 convolution layers. The ability of the proposed method to accurately localize speakers in a dynamic acoustic scenario with varying number of sources is also shown.

Daniele Mirabilii, Emanuël A. P. Habets

A novel multi-channel artificial wind noise generator based on a fluid dynamics model, namely the Corcos model, is proposed. In particular, the model is used to approximate the complex coherence function of wind noise signals measured with closely-spaced microphones in the free-field and for time-invariant wind stream direction and speed. Preliminary experiments focus on a spatial analysis of recorded wind noise signals and the validation of the Corcos model for diverse measurement set-ups. Subsequently, the Corcos model is used to synthetically generate wind noise signals exhibiting the desired complex coherence. The multi-channel generator is designed extending an existing single-channel generator to create N mutually uncorrelated signals, while the predefined complex coherence function is obtained exploiting an algorithm developed to generate multi-channel non-stationary noise signals under a complex coherence constraint. Temporal, spectral and spatial characteristics of synthetic signals match with those observed in measured wind noise. The artificial generation overcomes the time-consuming challenge of collecting pure wind noise samples for noise reduction evaluations and provides flexibility in the number of generated signals used in the simulations.

Fabian-Robert Stöter, Soumitro Chakrabarty, Bernd Edler et al.

The task of estimating the maximum number of concurrent speakers from single channel mixtures is important for various audio-based applications, such as blind source separation, speaker diarisation, audio surveillance or auditory scene classification. Building upon powerful machine learning methodology, we develop a Deep Neural Network (DNN) that estimates a speaker count. While DNNs efficiently map input representations to output targets, it remains unclear how to best handle the network output to infer integer source count estimates, as a discrete count estimate can either be tackled as a regression or a classification problem. In this paper, we investigate this important design decision and also address complementary parameter choices such as the input representation. We evaluate a state-of-the-art DNN audio model based on a Bi-directional Long Short-Term Memory network architecture for speaker count estimations. Through experimental evaluations aimed at identifying the best overall strategy for the task and show results for five seconds speech segments in mixtures of up to ten speakers.

Soumitro Chakrabarty, Emanuël A. P. Habets

The problem of multi-speaker localization is formulated as a multi-class multi-label classification problem, which is solved using a convolutional neural network (CNN) based source localization method. Utilizing the common assumption of disjoint speaker activities, we propose a novel method to train the CNN using synthesized noise signals. The proposed localization method is evaluated for two speakers and compared to a well-known steered response power method.

Dovid Y. Levin, Emanuël A. P. Habets, Sharon Gannot

Smartglasses, in addition to their visual-output capabilities, often contain acoustic sensors for receiving the user's voice. However, operation in noisy environments may lead to significant degradation of the received signal. To address this issue, we propose employing an acoustic sensor array which is mounted on the eyeglasses frames. The signals from the array are processed by an algorithm with the purpose of acquiring the user's desired near-filed speech signal while suppressing noise signals originating from the environment. The array is comprised of two AVSs which are located at the fore of the glasses' temples. Each AVS consists of four collocated subsensors: one pressure sensor (with an omnidirectional response) and three particle-velocity sensors (with dipole responses) oriented in mutually orthogonal directions. The array configuration is designed to boost the input power of the desired signal, and to ensure that the characteristics of the noise at the different channels are sufficiently diverse (lending towards more effective noise suppression). Since changes in the array's position correspond to the desired speaker's movement, the relative source-receiver position remains unchanged; hence, the need to track fluctuations of the steering vector is avoided. Conversely, the spatial statistics of the noise are subject to rapid and abrupt changes due to sudden movement and rotation of the user's head. Consequently, the algorithm must be capable of rapid adaptation. We propose an algorithm which incorporates detection of the desired speech in the time-frequency domain, and employs this information to adaptively update estimates of the noise statistics. Speech detection plays a key role in ensuring the quality of the output signal. We conduct controlled measurements of the array in noisy scenarios. The proposed algorithm preforms favorably with respect to conventional algorithms.