Several features computed from an audio signal have been shown to depend on the distance between the acoustic source and the receiver, but at the same time are heavily influenced by room characteristics and the microphone setup. While neural networks, if trained on signals representing a large variety of setups, have shown to deliver robust distance estimates from a coherent-to-diffuse power ratio (CDR) feature map at the input, we here push their modeling capabilities by additionally using the network as feature extractor. It is shown that distance estimation based on short-time Fourier transform (STFT) features can achieve a smaller estimation error and can operate on shorter signal segments compared to the previous CDR-based estimator.

Location: Auditorium

Binaural recording using artificial heads or microphones on or near the ears of a person, e.g., on modern headphones and hearables, is a well-established technology for spatial audio capture. Our previously presented Binaural Cue Adaptation method improves the reproduction of binaural recordings by adapting the signals recorded for a fixed head orientation to a listener’s dynamic head movements. In the present work, we exploit the fact that many binaural recording devices, such as hearables, are equipped with more than two microphones and generalize the method to use more than two microphone signals. Information from the additional microphone signals can be used to remove limitations on possible positions of sound sources in the recorded scene, and to improve the performance of the method in the presence of diffuse ambient signals and for larger head movements.

Location: Auditorium

Spatially distributed network nodes in Wireless Acoustic Sensor Networks (WASNs) offer different perspectives of an acoustic scene. While the spatial information contained in all recorded signals can in principle be exploited to determine the position of an acoustic Source of Interest (SoI), in challenging acoustic conditions, the observations provided by different network nodes are not equally reliable. Thus, accurately estimating the reliability of the nodes’ observations and incorporating them during the localization procedure is crucial, potentially even excluding observations. In this contribution, we demonstrate how a recently proposed microphone utility measure based on the correlation of single-channel feature streams can improve the localization accuracy when triangulating an acoustic source from node-wise Direction of Arrival (DOA) estimates.

Location: Auditorium

There are several approaches for acoustically localizing sound sources. This contribution utilizes an eight-channel microphone array that is mounted on an artificial head and spatially samples the direct vicinity of both ears. This allows for a localization approach that resembles the experience of a human listener while avoiding the limitations of an artificial head (e.g., with respect to front-back confusions).

The signals are used by a convolutional neural network - interpreting the localization task as a classification. The classification approach employs a novel use of binary classifiers that allows for implicitly taking the localization error into account. A performance assessment of the localization system is presented that is based on real audio recordings and a comparison with a related machine learning approach that uses a regular classification and two conventional beamforming algorithms is shown to highlight the positive impact of the new classifier design.

Location: P1 in Poster Area II

This paper takes a previously proposed convolutional recurrent deep neural network (DNN) approach to direction of arrival (DoA) estimation and extends this to perform 2D localisation using distributed microphone arrays. Triangulation on the individual DoAs from each array is the most straightforward extension of the original DNN. This paper proposes to allow more co-operation between the individual microphone arrays by sharing part of their neural network, in order to achieve a higher localisation accuracy. Two strategies will be discussed: one where the shared network has narrowband information, and one where only broadband information is shared. Robustness against slight clock offsets between different arrays is ensured by only sharing information at deeper layers in the DNN. The position and configuration of the microphone arrays are assumed known, in order to train the network. Simulations will show that combining information between neural network layers has a significant improvement over the triangulation approach.

Location: P2 in Poster Area II

Recently a method has been proposed to blindly estimate the geometry of an array of distributed microphones using reverberant speech, which relies on estimating the coherence matrix of the reverberation using an iterative expectation conditional-maximization (ECM) approach.Instead of using a data-independent initial estimate of the coherence matrix and a matched beamformer to estimate the initial speech and reverberation power spectral densities, in this paper we propose to use a data-dependent initial estimate of the coherence matrix and a time-varying minimum-power-distortionless-response beamformer. Simulation results show that the proposed ECM initialization significantly improves the estimation accuracy of the microphone array geometry and increases the generalizability for microphone arrays of different sizes.

Location: P3 in Poster Area II

In this paper, we propose and evaluate a method for binaural speaker localization using modulation-domain features extracted from the binaural microphone signals of hearing devices. In contrast to most other localization methods the proposed method does not require perfectly synchronized audio signals from the left and the right ear but uses front/back cardioid signals and a classification approach with a small number (< 50) of modulation-domain features for each signal frame. The method employs an efficient implementation of such features using a bank of recursive IIR filters which makes it suitable for low-power portable devices and also allows the application of data-driven optimization procedures. We analyze the capability of these features to reflect not only interaural level differences but also temporal modulation patterns. We evaluate our method on simulated and real-world binaural signals and compare the proposed approach to a beamforming-based method which requires fully-synchronized microphone signals.

Location: P4 in Poster Area II

Despite their small share in overall signal energy, plosives have been previously shown to be important for speech perception. We propose a simple, yet effective, model-based phase-aware speech enhancement approach specifically targeted at plosives. Starting from a model of the plosive burst as a unit impulse, we introduce three phase enhancement schemes: simple replacement of the noisy phase with a linear function, linear regression, as well as smoothing by local polynomial regression. To improve the outcome and compensate for model mismatch we also propose an SNR-based weighting. All schemes are evaluated under both oracle and realistic conditions, showing a consistent improvement in instrumentally predicted speech quality and, to a lesser degree, speech intelligibility. When only frames containing plosives are considered, a segmental SNR improvement of 2 dB to 6 dB can be observed, depending on the input SNR.

Location: Auditorium

Speaker-conditioned target speaker extraction systems rely on auxiliary information about the target speaker to extract the target speaker signal from a mixture of multiple speakers. Typically, a deep neural network is applied to isolate the relevant target speaker characteristics. In this paper, we focus on a single-channel target speaker extraction system based on a CNN-LSTM separator network and a speaker embedder network requiring reference speech of the target speaker. In the LSTM layer of the separator network, we propose to customize the LSTM cells in order to only remember the specific voice patterns corresponding to the target speaker by modifying the information processing in the forget gate. Experimental results for two-speaker mixtures using the Librispeech dataset show that this customization significantly improves the target speaker extraction performance compared to using standard LSTM cells.

Location: Auditorium

In some cases, a speech communication link is necessary in underwater environments. While in terms of data communication advanced techniques are investigated and employed, mostly traditional (analogue) forms of speech transmission are used for speech communication (at least in commercial products). In this contribution we present a mixed analog-digital approach, that tackles this problem by combining digital and analogue approaches. Both transmission paths rely on each other in order to lower the necessary bitrate for the digital part and enhance the signal properties of the analogue part. The proposed transmission scheme uses linear predictive coding (LPC) filtering combined with a codebook to whiten the input speech. Afterwards, the resulting signal is power normalized and transmitted in the analog transmission part and the quantized normalization factor is transmitted together with the codebook index in the digital transmission part. This leads to an analogue transmission signal which is strongly reduced in dynamics.

Location: Auditorium

Permutation invariant training (PIT) is a widely used training criterion for neural network-based source separation, used for both utterance-level separation with utterance-level PIT (uPIT) and separation of long recordings with the recently proposed Graph-PIT. When implemented naively, both suffer from an exponential complexity in the number of utterances to separate, rendering them unusable for large numbers of speakers or long realistic recordings. We present a decomposition of the PIT criterion into the computation of a matrix and a strictly monotonously increasing function so that the permutation or assignment problem can be solved efficiently with several search algorithms. The Hungarian algorithm can be used for uPIT and we introduce various algorithms for the Graph-PIT assignment problem to reduce the complexity to be polynomial in the number of utterances.

Location: P1 in Poster Area III

In this paper, we compare two extended generalized sidelobe canceller (GSC) structures, which exploit external microphones in conjunction with a local microphone array to improve the noise and interferer reduction. As a baseline algorithm we consider a local GSC using only the local microphones, for which the relative transfer function (RTF) vector of the target speaker is known. To incorporate the external microphones in a minimum power distortionless response beamformer, the RTF vector of the target speaker needs to be estimated. Since the estimation accuracy of this RTF vector depends on the signal-to-interferer ratio, the GSC with external speech references (GSC-ESR) pre-processes the external microphone signals to reduce the interferer. In a simplified extended structure, namely the GSC with external references (GSC-ER) no such pre-filtering operation is performed. Simulation results show that the GSC-ESR structure yields the best results in terms of noise and interferer reduction, especially in adverse conditions.

Location: P2 in Poster Area III

In IP-based teleconferences disturbing howling artifacts might occur whenever multiple participants are in the same room using different communication devices. In this paper it is described how the feedback problem can be reformulated as a system identification task and what are the specifics compared to the classical field of feedback cancellation. The deployment of a Kalman filter based adaptation in the frequency domain is proposed due to its robustness against long phases of double talk and near-end single talk. In order to account for sudden changes in the network delay or in the feedback path a shadow filter system is additionally introduced. Simulation results confirm that the proposed system performs an effective feedback cancellation improving the communication quality significantly.

Location: P3 in Poster Area III

We consider the problem of simultaneous reduction of ego-noise, i.e., the noise produced by a robot, and environmental noise. Both noise types may occur simultaneously for humanoid interactive robots. Dictionary- and template-based approaches have been proposed for ego-noise reduction. However, most of them lack adaptability to unseen noise types and thus exhibit limited performance in real-world scenarios with environmental noise. Recently, a variational autoencoder (VAE)-based speech model combined with a fully-adaptive dictionary-based noise model, i.e., non-negative matrix factorization (NMF), has been proposed for environmental noise reduction, showing decent adaptability to unseen noise data. In this paper, we propose to extend this framework with a partially-adaptive dictionary-based noise model, which partly adapts to unseen environmental noise while keeping the part pre-trained on ego-noise unchanged. With appropriate sizes, we demonstrate that the partially-adaptive approach outperforms the approaches based on the fully-adaptive and completely-fixed dictionaries, respectively.

Location: P4 in Poster Area III

Due to the increasing availability of devices with multiple microphones, modern acoustic echo cancellation systems need to estimate and optimize a large number of adaptive filters, resulting in an ever-increasing computational cost of such systems. In this paper, a highly efficient multichannel system identification method is introduced for setups where multiple microphone signals are fused using beamformers. By exploiting the knowledge on these beamformers, the proposed approach achieves high levels of computational efficiency, similar to those of beamformer-first acoustic echo cancellation methods, while avoiding severe performance drops resulting from the use of time-varying beamformers. This is validated in an experimental part using both, echo cancellation and system identification performance metrics.

Location: P5 in Poster Area III

Applying multichannel noise reduction in the spherical harmonic domain (SHD) can be computationally expensive. To reduce the complexity, an approach is proposed that combines signal-dependent beamforming with spatial warping for dimensionality reduction. In particular, the multichannel Wiener filter combined with spatial warping is investigated. Two different methods to combine beamforming with warping are investigated, both aiming to reduce the computational complexity of the multichannel Wiener filter, which can be mainly attributed to the power spectral density matrix inversions required to compute the beamforming weights. A performance evaluation using simulated reverberant spherical microphone array signals is conducted wherein it is shown that the proposed methods significantly reduce the computational complexity whilst providing high degrees of signal enhancement when compared to order-truncation of the SHD signal without warping.

Location: P6 in Poster Area III

Unsupervised blind source separation methods do not require a training phase and thus cannot suffer from a train-test mismatch, which is a common concern in neural network based source separation. The unsupervised techniques can be categorized in two classes, those building upon the sparsity of speech in the Short-Time Fourier transform domain and those exploiting non-Gaussianity or non-stationarity of the source signals. In this contribution, spatial mixture models which fall in the first category and independent vector analysis (IVA) as a representative of the second category are compared w.r.t. their separation performance and the performance of a downstream speech recognizer on a reverberant dataset of reasonable size. Furthermore, we introduce a serial concatenation of the two, where the result of the mixture model serves as initialization of IVA, which achieves significantly better WER performance than each algorithm individually and even approaches the performance of a complex neural network based technique.

Location: P7 in Poster Area III

Acoustic feedback compensation in speech applications often suffers from an in-sufficient adaptation. Aa a result feedback whistling may be audible and the compensation gain and generally the speech quality is limited. In this work,we introduce the concept of an energy-decay operator and show how to use it for tap-selective step size control of a multi-delay frequency domain filter (MDF). Therefore, we derive a so-called "complex sign-sign algorithm" with an inherently included decay operator. Finally, we show robust simulation results with high amplification although the used algorithm shows very low complexity. Possible applications are in-car communication (ICC) systems or other sound reinforcement applications.

Location: P8 in Poster Area III

This paper proposes to unify two deep-learning methods, CountNet and Deep Clustering, designed for speaker count and separation respectively, in order to perform speaker count agnostic speech separation. Two approaches are compared, where the speaker count estimation and separation subnetworks are either trained separately or jointly. Training and evaluation are conducted on a tailored dataset WSJ0-mixN, which is an extension of the WSJ0-mix2 and WSJ0-mix3 datasets for an arbitrary number of speakers.Results show that both systems are capable of separating up to four sources without prior information on the number of speakers. Furthermore, the joint approach is able to perform similarly to its separate counterpart while using 46% fewer parameters.

Location: P9 in Poster Area III

Recently, the convolutional weighted power minimization distortionless response (WPD) beamformer was proposed, which unifies multi-channel weighted prediction error dereverberation and minimum power distortionless response beamforming. To optimize the convolutional filter, the desired speech component is modeled with a time-varying Gaussian model, which promotes the sparsity of the desired speech component in the short-time Fourier transform domain compared to the noisy microphone signals. In this paper we generalize the convolutional WPD beamformer by using an lp-norm cost function, introducing an adjustable shape parameter which enables to control the sparsity of the desired speech component. Experiments based on the REVERB challenge dataset show that the proposed method outperforms the conventional convolutional WPD beamformer in terms of objective speech quality metrics.

Location: P10 in Poster Area III

Wireless Acoustic Sensor Networks (WASNs) have a wide range of audio signal processing applications. Due to the spatial diversity of the microphone and their relative position to the acoustic source, not all microphones are equally useful for subsequent audio signal processing tasks, nor do they all have the same wireless data transmission rates. Hence, a central task in WASNs is to balance a microphone’s estimated acoustic utility against its transmission delay, selecting a best-possible subset of microphones to record audio signals.In this work, we use reinforcement learning to decide if a microphone should be used or switched off to maximize the acoustic quality at low transmission delays, while minimizing switching frequency. In experiments with moving sources in a simulated acoustic environment, our method outperforms naive baseline comparisons.

Location: P11 in Poster Area III

Development of pathological speech systems is currently hindered by the lack of a standardised objective evaluation framework. In this work, (1) we utilise existing detection andanalysis techniques to propose a general framework for the consistent evaluation of synthetic pathological speech, then (2)using our proposed framework, we develop a dysarthric voice conversion system (VC) using CycleGAN-VC, and show that the developed system is able to synthesise dysarthric speech with different levels of speech intelligibility.

Location: Auditorium

Recently proposed automatic pathological speech classification techniques use unsupervised auto-encoders to obtain a high-level abstract representation of speech. Since these representations are learned based on reconstructing the input, there is no guarantee that they are robust to pathology-unrelated cues such as speaker identity information. Further, these representations are not necessarily discriminative for pathology detection. In this paper, we exploit supervised auto-encoders to extract robust and discriminative speech representations for Parkinson's disease classification. To reduce the influence of speaker variabilities unrelated to pathology, we propose to obtain speaker identity-invariant representations by adversarial training of an auto-encoder and a speaker identification task. To obtain a discriminative representation, we propose to jointly train an auto-encoder and a pathological speech classifier. Experimental results on a Spanish database show that the proposed supervised representation learning methods yield more robust and discriminative representations for automatically classifying Parkinson's disease speech, outperforming the baseline unsupervised representation learning system.

Location: Auditorium

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.

Location: Auditorium

In the human cortex, the auditory signal is segmented ro-bustly into syllables using θ-oscillations, where the phase and instantaneous frequency of each θ-cycle corresponds to the position and duration of a syllable. Recently, in the superior temporal cortex, ensamples of neurons sensitive to edge features have been detected, which spike at the maximal rise of the envelope of the auditory signal at the onset of the nucleus of a syllable. The paper presents a cortical model of an θ-oscillator driven by features of the envelope of the auditory signal and driven by edge features resetting the θ-phase. First experiments on the performance of the θ-oscillator are presented.

Location: P1 in Poster Area IV

In times of rising noise pollution, the demand for active noise control (ANC) in hearable devices continues to increase rapidly. To date, time-invariant feedforward systems constitute a favorable option for their implementation. Recently, active acoustic equalization (AAE) was proposed as a generalization to the feedforward design problem. It facilitates the design of digital filters for various use-cases including ANC- and hear-through-applications. In this contribution, we analyze performance bounds of time-invariant AAE systems and discuss their implications. We build upon techniques which are established in the field of ANC to provide estimates for arbitrary use-cases. The approach facilitates an isolated analysis of individual aspects such as transfer path variance and processing delay.

Location: P2 in Poster Area IV

This study is concerned with the relation between the information-theoretic notion of surprisal and articulatory gesture in Polish consonant-to-vowel transitions. It addresses the question of the influence of diphone predictability on spectral trajectories and articulatory gestures by relating the effect of surprisal with motor fluency. The study triangulates the computation of locus equations (LE) with kinematic data obtained from electromagnetic articulograph (EMA). The kinematic and acoustic data showed that a small coarticulation effect was present in the high- and low-surprisal clusters. Regardless of some small discrepancies across the measures, a high degree of overlap of adjacent segments is reported for the mid-surprisal group in both domains. Two explanations of the observed effect are proposed. The first refers to low-surprisal coarticulation resistance and suggests the need to disambiguate predictable sequences. The second, observed in high surprisal clusters, refers to the prominence given to emphasize the unexpected concatenation.

Location: P3 in Poster Area IV

State-of-the-art drone detection systems are generally combining different sensors, such as radar, acoustic, radio frequency (RF-) or optic sensors, each of which contributes individual capabilities. Acoustic sensors have a relatively low spatial range. On the other hand, they can be used in conditions of bad visibility and for scenarios without line-of-sight between sensor and drone. The development of acoustic drone detection (ADD) algorithms typically requires substantial amounts of realistic recordings of flying drones. Moreover, when evaluating methods for robustly extracting characteristic properties of drones, such as the fundamental frequency (F0) of the rotor blades from sensor data, knowledge of ground truth (GT) data is important. In this paper we present a framework for automatically generating both acoustic drone recordings and GT transcripts of the corresponding time-synchronous motor rotations. We present an application using the thus generated data for evaluating an ADD algorithm based on F0 tracking.

Location: P4 in Poster Area IV

Conversations over a telephone network require a timely transmission of speech to enable smooth interaction. When a transmission delay is introduced, turn-taking signals arrive too late, and the conversational quality degrades. This disruption of the conversation flow also depends on the interactivity of the conversation. Current instrumental quality models do not take into account the interactivity of a conversation. However, the simulation of conversations has proven to replicate the turn-taking behavior of conversations with different interactivity levels. It can also model the changes in interactivity when transmission delay is introduced. In this paper, we simulate two types of conversations at various levels of transmission delay. We perform a parametric conversation analysis to extract interactivity parameters and use them to predict the conversational quality. We compare the results to the predictions of the E-model, an instrumental transmission planning model, and quality ratings obtained in a conversation experiment.

Location: Auditorium

Speech communication and dialog systems inside vehicles are usually optimized for a speaker in the driver’s seat who is permanently facing forward. However while driving, a speaker does repeatedly rotate his head left and right, which changes the properties of the signal picked up by the microphone. This work analyzes to what extend such a speaker head rotation impacts the signal in terms of frequency content, listening quality, and listening effort. By presenting a simple method for compensating the rotation effects and by combining this method with noise reduction and band limitation of the signals, an analysis of the far-end listening situation with and without speaker head rotation compensation is achieved.

Location: Auditorium

Four non-intrusive models are compared that predict human speech recognition thresholds (SRTs, i.e., signal to noise ratios with 50% word recognition rate) in different acoustic environments. Three of them use the blind binaural processing stage (bBSIM) as front-end, while one model uses the spectral representation of the left and right ear signal channels together with their difference. Predictions are evaluated for three acoustic environments (anechoic, office, and cafeteria) with speech from the front and noise from different directions. Despite many technical differences across the models, all of them perform quite accurately (root mean squared prediction errors below 2.2 dB for all models). This implies that any of the non-intrusive models facilitates to predict SRTs for listeners with normal hearing measured in stationary noise, different acoustic environments, and spatial configurations.

Location: Auditorium

In this paper we present an open database for the development of detection and enhancement algorithms of speech transmitted over HF radio channels. It consists of audio samples recorded by various receivers at different locations across Europe, all monitoring the same single-sideband modulated transmission from a base station in Paderborn, Germany. Transmitted and received speech signals are precisely time aligned to offer parallel data for supervised training of deep learning based detection and enhancement algorithms. For the task of speech activity detection two exemplary baseline systems are presented, one based on statistical methods employing a multi-stage Wiener filter with minimum statistics noise floor estimation, and the other relying on a deep learning approach.

Location: P1 in Poster Area V

While Fourier phase has long been considered unimportant for speech enhancement in the short-time Fourier domain, phase-aware speech processing is receiving increasing attention in recent years. Among other advances, it has been shown that when using very short frames (<2 ms), the phase spectrum carries enough information to allow an intelligible reconstruction from phase alone, whereas the magnitude spectrum remains virtually bare of any information relevant to intelligibility. Formal listening experiments are expensive and laborious. Hence, in order to facilitate further research into very short frames, researchers require reliable instrumental intelligibility measures. We present a study of two common measures: Short-Time Objective Intelligibility (STOI) and Extended STOI (ESTOI) and compare their performance for speech reconstruction from phase. Our results indicate that only ESTOI is able to predict the general trend of intelligibility across a wide range of frame lengths (including very short frames) for both phase-based and magnitude-based reconstructions.

Location: P2 in Poster Area V

A reproducible acoustic environment forms the basis for scientific evaluations of speech enhancement systems, as well as their influence on speech quality and intelligibility. Studies with people usually use binaural recordings played back through headphones. However, the effort required to evaluate such a speech enhancement system in this way can quickly become impractical, especially if the system is a black box and does not allow access to internal signals. In order to be able to investigate such systems and the associated influences on the listeners, a reproducible acoustic environment simulation has been developed that is able to generate different noise scenarios in a soundproof room. The system uses loudspeakers to place both the listeners and the communication system in a realistic noise scenario. It recreates the statistical properties of the desired sound scenario at corresponding microphones using real recordings as reference.

Location: P3 in Poster Area V

Speech communication under adverse conditions may be extremely stressful for the person located at the receiving side. There, background noise may originate from the environment and cannot be reduced for the listener. Additionally, far-end speech might be degraded by the network, due to e.g., transcoding or packet loss. The assessment of several issues mentioned above is partly already addressed by existing state-of-the-art methods. They are often related to speech quality, though. Examples include ETSI103281 for noise reduction in send direction or ITU T P863 for network related processing. However, when quality is already considerably degraded, an intelligibility related measure is of higher interest. As an alternative to the assessment of intelligibility, perceived listening effort was found to be a more suitable measure for such scenarios in several recent studies. This contribution presents the latest developments of an instrumental prediction model that is based on a comprehensive amount of subjective databases.

Location: P4 in Poster Area V