Szymon Drgas
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Szymon Drgas has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
- Adam Dąbrowski
Poznan University of Technology, Faculty of Control, Robotics & Electrical Engineering, Institute of Automatic Control and Robotics, ul. Piotrowo 3a, 60-965 Poznan, Poland
3 shared publications - Tomasz Grzywalski
Department of Information Technology, Ghent University, Belgium
3 shared publications - Tuomas Virtanen
Faculty of Information Technology and Communication Sciences, Tampere University, Tampere, Finland
2 shared publications - Anna Przekoracka-Krawczyk
Laboratory of Vision Science and Optometry, Faculty of Physics and Astronomy, Adam Mickiewicz University in Poznań, Poznań, Poland
1 shared publication Antti Hurmalainen
Yousician Ltd., Helsinki, Finland
1 shared publication- Archontis Politis
Tampere University
1 shared publication
Latest Publications
- 5 April 2024
- Published by Elsevier in Speech Communication
- Vol. 159
Speech intelligibility prediction using generalized ESTOI with fine-tuned parameters
- 5 April 2024
- Published by Elsevier in Speech Communication
- Vol. 159
Speech intelligibility prediction using generalized ESTOI with fine-tuned parameters
- 30 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 32 (23299290) , 203-214
Dynamic Processing Neural Network Architecture for Hearing Loss Compensation
This paper proposes neural networks for compensating sensorineural hearing loss. The aim of the hearing loss compensation task is to transform a speech signal to increase speech intelligibility after further processing by a person with a hearing impairment, which is modeled by a hearing loss model. We propose an interpretable model called dynamic processing network, which has a structure similar to band-wise dynamic compressor. The network is differentiable, and therefore allows to learn its parameters to maximize speech intelligibility. More generic models based on convolutional layers were tested as well. The performance of the tested architectures was assessed using spectro-temporal objective index (STOI) with hearing-threshold noise and hearing aid speech intelligibility (HASPI) metrics. The dynamic processing network gave a significant improvement of STOI and HASPI in comparison to popular compressive gain prescription rule Camfit. A large enough convolutional network could outperform the interpretable model with the cost of larger computational load. Finally, a combination of the dynamic processing network with convolutional neural network gave the best results in terms of STOI and HASPI.
- 30 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 32 (23299290) , 203-214
Dynamic Processing Neural Network Architecture for Hearing Loss Compensation
This paper proposes neural networks for compensating sensorineural hearing loss. The aim of the hearing loss compensation task is to transform a speech signal to increase speech intelligibility after further processing by a person with a hearing impairment, which is modeled by a hearing loss model. We propose an interpretable model called dynamic processing network, which has a structure similar to band-wise dynamic compressor. The network is differentiable, and therefore allows to learn its parameters to maximize speech intelligibility. More generic models based on convolutional layers were tested as well. The performance of the tested architectures was assessed using spectro-temporal objective index (STOI) with hearing-threshold noise and hearing aid speech intelligibility (HASPI) metrics. The dynamic processing network gave a significant improvement of STOI and HASPI in comparison to popular compressive gain prescription rule Camfit. A large enough convolutional network could outperform the interpretable model with the cost of larger computational load. Finally, a combination of the dynamic processing network with convolutional neural network gave the best results in terms of STOI and HASPI.
- Open Access
A Survey on Low-Latency DNN-Based Speech Enhancement
This paper presents recent advances in low-latency, single-channel, deep neural network-based speech enhancement systems. The sources of latency and their acceptable values in different applications are described. This is followed by an analysis of the constraints imposed on neural network architectures. Specifically, the causal units used in deep neural networks are presented and discussed in the context of their properties, such as the number of parameters, the receptive field, and computational complexity. This is followed by a discussion of techniques used to reduce the computational complexity and memory requirements of the neural networks used in this task. Finally, the techniques used by the winners of the latest speech enhancement challenges (DNS, Clarity) are shown and compared.
Open AccessA Survey on Low-Latency DNN-Based Speech Enhancement
This paper presents recent advances in low-latency, single-channel, deep neural network-based speech enhancement systems. The sources of latency and their acceptable values in different applications are described. This is followed by an analysis of the constraints imposed on neural network architectures. Specifically, the causal units used in deep neural networks are presented and discussed in the context of their properties, such as the number of parameters, the receptive field, and computational complexity. This is followed by a discussion of techniques used to reduce the computational complexity and memory requirements of the neural networks used in this task. Finally, the techniques used by the winners of the latest speech enhancement challenges (DNS, Clarity) are shown and compared.
- Open Access
Speech Enhancement by Multiple Propagation through the Same Neural Network
Monaural speech enhancement aims to remove background noise from an audio recording containing speech in order to improve its clarity and intelligibility. Currently, the most successful solutions for speech enhancement use deep neural networks. In a typical setting, such neural networks process the noisy input signal once and produces a single enhanced signal. However, it was recently shown that a U-Net-based network can be trained in such a way that allows it to process the same input signal multiple times in order to enhance the speech even further. Unfortunately, this was tested only for two-iteration enhancement. In the current research, we extend previous efforts and demonstrate how the multi-forward-pass speech enhancement can be successfully applied to other architectures, namely the ResBLSTM and Transformer-Net. Moreover, we test the three architectures with up to five iterations, thus identifying the method’s limit in terms of performance gain. In our experiments, we used the audio samples from the WSJ0, Noisex-92, and DCASE datasets and measured speech enhancement quality using SI-SDR, STOI, and PESQ. The results show that performing speech enhancement up to five times still brings improvements to speech intelligibility, but the gain becomes smaller with each iteration. Nevertheless, performing five iterations instead of two gives additional a 0.6 dB SI-SDR and four-percentage-point STOI gain. However, these increments are not equal between different architectures, and the U-Net and Transformer-Net benefit more from multi-forward pass compared to ResBLSTM.
Open AccessSpeech Enhancement by Multiple Propagation through the Same Neural Network
Monaural speech enhancement aims to remove background noise from an audio recording containing speech in order to improve its clarity and intelligibility. Currently, the most successful solutions for speech enhancement use deep neural networks. In a typical setting, such neural networks process the noisy input signal once and produces a single enhanced signal. However, it was recently shown that a U-Net-based network can be trained in such a way that allows it to process the same input signal multiple times in order to enhance the speech even further. Unfortunately, this was tested only for two-iteration enhancement. In the current research, we extend previous efforts and demonstrate how the multi-forward-pass speech enhancement can be successfully applied to other architectures, namely the ResBLSTM and Transformer-Net. Moreover, we test the three architectures with up to five iterations, thus identifying the method’s limit in terms of performance gain. In our experiments, we used the audio samples from the WSJ0, Noisex-92, and DCASE datasets and measured speech enhancement quality using SI-SDR, STOI, and PESQ. The results show that performing speech enhancement up to five times still brings improvements to speech intelligibility, but the gain becomes smaller with each iteration. Nevertheless, performing five iterations instead of two gives additional a 0.6 dB SI-SDR and four-percentage-point STOI gain. However, these increments are not equal between different architectures, and the U-Net and Transformer-Net benefit more from multi-forward pass compared to ResBLSTM.
- 9 March 2022
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 81 (13) , 18617-18639
Speech enhancement using U-nets with wide-context units
- 9 March 2022
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 81 (13) , 18617-18639
Speech enhancement using U-nets with wide-context units
- 14 September 2021
- Published by American Speech Language Hearing Association in Journal of Speech, Language, and Hearing Research
- Vol. 64 (9) , 3603-3616
The Combination of Neural Tracking and Alpha Power Lateralization for Auditory Attention Detection
Purpose: The acoustic source that is attended to by the listener in a mixture can be identified with a certain accuracy on the basis of their neural response recorded during listening, and various phenomena may be used to detect attention. For example, neural tracking (NT) and alpha power lateralization (APL) may be utilized in order to obtain information concerning attention. However, these methods of auditory attention detection (AAD) are typically tested in different experimental setups, which makes it impossible to compare their accuracy. The aim of this study is to compare the accuracy of AAD based on NT, APL, and their combination for a dichotic natural speech listening task. Method: Thirteen adult listeners were presented with dichotic speech stimuli and instructed to attend to one of them. Electroencephalogram of the subjects was continuously recorded during the experiment using a set of 32 active electrodes. The accuracy of AAD was evaluated for trial lengths of 50, 25, and 12.5 s. AAD was tested for various parameters of NT- and APL-based modules. Results: The obtained results suggest that NT of natural running speech provides similar accuracy to APL. The statistically significant improvement of the accuracy of AAD using a combined method has been observed not only for the longest duration of test samples (50 s, p = .005) but also for shorter ones (25 s, p = .011). Conclusions: It seems that the combination of standard NT and APL significantly increases the effectiveness of accurate identification of the traced signal perceived by a listener under dichotic conditions. It has been demonstrated that, under certain conditions, the combination of NT and APL may provide a benefit for AAD in cocktail party scenarios.
- 14 September 2021
- Published by American Speech Language Hearing Association in Journal of Speech, Language, and Hearing Research
- Vol. 64 (9) , 3603-3616
The Combination of Neural Tracking and Alpha Power Lateralization for Auditory Attention Detection
Purpose: The acoustic source that is attended to by the listener in a mixture can be identified with a certain accuracy on the basis of their neural response recorded during listening, and various phenomena may be used to detect attention. For example, neural tracking (NT) and alpha power lateralization (APL) may be utilized in order to obtain information concerning attention. However, these methods of auditory attention detection (AAD) are typically tested in different experimental setups, which makes it impossible to compare their accuracy. The aim of this study is to compare the accuracy of AAD based on NT, APL, and their combination for a dichotic natural speech listening task. Method: Thirteen adult listeners were presented with dichotic speech stimuli and instructed to attend to one of them. Electroencephalogram of the subjects was continuously recorded during the experiment using a set of 32 active electrodes. The accuracy of AAD was evaluated for trial lengths of 50, 25, and 12.5 s. AAD was tested for various parameters of NT- and APL-based modules. Results: The obtained results suggest that NT of natural running speech provides similar accuracy to APL. The statistically significant improvement of the accuracy of AAD using a combined method has been observed not only for the longest duration of test samples (50 s, p = .005) but also for shorter ones (25 s, p = .011). Conclusions: It seems that the combination of standard NT and APL significantly increases the effectiveness of accurate identification of the traced signal perceived by a listener under dichotic conditions. It has been demonstrated that, under certain conditions, the combination of NT and APL may provide a benefit for AAD in cocktail party scenarios.
- 23 September 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 157-162
Speech enhancement by iterating forward pass through U-net
- 23 September 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 157-162
Speech enhancement by iterating forward pass through U-net
- 29 May 2017
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 25 (8) , 1644-1656
Binary Non-Negative Matrix Deconvolution for Audio Dictionary Learning
In this study, we propose an unsupervised method for dictionary learning in audio signals. The new method, called binary nonnegative matrix deconvolution (BNMD), is developed and used to discover patterns from magnitude-scale spectrograms. The BNMD models an audio spectrogram as a sum of delayed patterns having binary gains (activations). Only small subsets of patterns can be active for a given spectrogram excerpt. The proposed method was applied to speaker identification and separation tasks. The experimental results show that dictionaries obtained by the BNMD bring much higher speaker identification accuracies averaged over a range of SNRs from -6 dB to 9 dB (91.3%) than the NMD-based dictionaries (37.8-75.4%). The BNMD also gives a benefit over dictionaries obtained using vector quantization (87.8%). For bigger dictionaries the difference between the BNMD and the vector quantization (VQ) is getting smaller. For the speech separation task the BNMD dictionary gave a slight improvement over the VQ.
- 29 May 2017
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 25 (8) , 1644-1656
Binary Non-Negative Matrix Deconvolution for Audio Dictionary Learning
In this study, we propose an unsupervised method for dictionary learning in audio signals. The new method, called binary nonnegative matrix deconvolution (BNMD), is developed and used to discover patterns from magnitude-scale spectrograms. The BNMD models an audio spectrogram as a sum of delayed patterns having binary gains (activations). Only small subsets of patterns can be active for a given spectrogram excerpt. The proposed method was applied to speaker identification and separation tasks. The experimental results show that dictionaries obtained by the BNMD bring much higher speaker identification accuracies averaged over a range of SNRs from -6 dB to 9 dB (91.3%) than the NMD-based dictionaries (37.8-75.4%). The BNMD also gives a benefit over dictionaries obtained using vector quantization (87.8%). For bigger dictionaries the difference between the BNMD and the vector quantization (VQ) is getting smaller. For the speech separation task the BNMD dictionary gave a slight improvement over the VQ.
- 26 September 2016
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 811-815
Allophones in automatic whispery speech recognition
The article presents studies on the automatic whispery speech recognition. In the performed research a new corpus with whispery speech has been used. It has been checked whether the extended set of articulatory units (allophones have been used instead of phonemes) improves quality of whispery speech recognition. Experimental results show that the small changes in the allophone set may provide better speech recognition quality than using phonemes approach. The authors also made available the trained g2p (grapheme-to-phoneme) model of Polish language for the Sequitur toolkit.
- 26 September 2016
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 811-815
Allophones in automatic whispery speech recognition
The article presents studies on the automatic whispery speech recognition. In the performed research a new corpus with whispery speech has been used. It has been checked whether the extended set of articulatory units (allophones have been used instead of phonemes) improves quality of whispery speech recognition. Experimental results show that the small changes in the allophone set may provide better speech recognition quality than using phonemes approach. The authors also made available the trained g2p (grapheme-to-phoneme) model of Polish language for the Sequitur toolkit.
- 15 August 2015
- Published by Springer Nature
- p. 462-469
Speaker Verification Using Adaptive Dictionaries in Non-negative Spectrogram Deconvolution
- 15 August 2015
- Published by Springer Nature
- p. 462-469
Speaker Verification Using Adaptive Dictionaries in Non-negative Spectrogram Deconvolution