About this Research Topic
In practice, a desired acoustic source usually cannot be observed alone but only superimposed by interferers and noise. Source separation and extraction is an important task in audio signal processing, attracting much research interest. Many of these approaches make only very generic assumptions about the acoustic environment like statistical independence of the source signals (e.g., blind source separation), the direction of arrival of the target source (e.g., beamforming), or spectral/waveform properties of the signal to be extracted (e.g., DNN-based approaches). However, often some additional side information is available or could be generated by additional algorithms. These may support the signal extraction/separation algorithm to cope with uncertainties, increase performance and robustness, add new capabilities, etc.
This Research Topic addresses research related to acoustic signal separation and extraction relying on side information about the acoustic scene. This may include new signal processing algorithms that make use of: information that is available in addition to the observed microphone signals in certain relevant use cases; approaches that rely on the estimation of side information for acoustic signal separation and extraction; methods exploiting multiple signal modalities; hybrid methods combining physics-inspired or information-theory-based ideas with learning from examples, etc. The proposed algorithms may be based on classical signal processing (e.g., beamforming), statistical signal processing (e.g., blind source separation techniques), or machine learning-based techniques (e.g., based on DNNs). Use cases may include hands-free human machine interfaces, hearing aids, smart devices, automatic speech recognition systems.
Contributions to this Research Topic include but are not limited to the following list of research fields:
• Algorithms estimating side information for informed signal separation and extraction,
• Robust automatic speech recognition relying on speaker extraction, in particular methods:
o relying on side information such as speaker diarization and identification,
o trained end-to-end, such that the requirements of the ASR system inform the speaker extraction stage,
• Classical signal processing and array signal processing potentially combined with machine learning approaches:
o Innovative approaches for beamforming relying on side information, e.g., for selecting the target source, for selecting an expert beamformer out of a set of available beamformers, etc.
• Neural beamforming:
o Innovative methods for estimating signal statistics relying on side information for signal-dependent beamformers.
• Blind source separation and extraction (BSS/BSE):
o BSS/BSE methods making use of spatial information about the target source,
o BSS/BSE methods relying on specialised source models, e.g., learned source models or statistical models for specific application scenarios.
• Deep learning-based approaches:
o DNN-based methods that rely on side information for selecting a target source or adapting signal models, e.g., one-shot speaker adaptation.
• Hybrid methods:
o Methods combining physical and information-theoretical models with learning-based structures in any way.
This Research Topic addresses research related to acoustic signal separation and extraction relying on side information about the acoustic scene. This may include new signal processing algorithms that make use of: information that is available in addition to the observed microphone signals in certain relevant use cases; approaches that rely on the estimation of side information for acoustic signal separation and extraction; methods exploiting multiple signal modalities; hybrid methods combining physics-inspired or information-theory-based ideas with learning from examples, etc. The proposed algorithms may be based on classical signal processing (e.g., beamforming), statistical signal processing (e.g., blind source separation techniques), or machine learning-based techniques (e.g., based on DNNs). Use cases may include hands-free human machine interfaces, hearing aids, smart devices, automatic speech recognition systems.
Contributions to this Research Topic include but are not limited to the following list of research fields:
• Algorithms estimating side information for informed signal separation and extraction,
• Robust automatic speech recognition relying on speaker extraction, in particular methods:
o relying on side information such as speaker diarization and identification,
o trained end-to-end, such that the requirements of the ASR system inform the speaker extraction stage,
• Classical signal processing and array signal processing potentially combined with machine learning approaches:
o Innovative approaches for beamforming relying on side information, e.g., for selecting the target source, for selecting an expert beamformer out of a set of available beamformers, etc.
• Neural beamforming:
o Innovative methods for estimating signal statistics relying on side information for signal-dependent beamformers.
• Blind source separation and extraction (BSS/BSE):
o BSS/BSE methods making use of spatial information about the target source,
o BSS/BSE methods relying on specialised source models, e.g., learned source models or statistical models for specific application scenarios.
• Deep learning-based approaches:
o DNN-based methods that rely on side information for selecting a target source or adapting signal models, e.g., one-shot speaker adaptation.
• Hybrid methods:
o Methods combining physical and information-theoretical models with learning-based structures in any way.
Keywords: Audio Source Separation, Algorithms, Acoustic signal extraction, Signal modalities, Machine-learning
Important Note: All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.
