Perceptually motivated restoration of highly degraded audio signals
Highly degraded audio restoration presents a live research field currently. However, the
recovery of missing, saturated or quantized audio signal information has been studied only
marginally in connection to psychoacoustics. The project will provide new models and
techniques for audio signal restoration and compression, resulting in significantly improved
perceptual quality of the restored audio, compared to the state-of-the-art approaches. The new
techniques will be based on convex and non-convex optimization and they will be encompassed
by a unifying theoretical framework. The quality of restoration will be evaluated using both the
perceptually-motivated objective measures and formal listening tests. The experimental part will
answer the question whether simultaneous storage of subsampled and quantized data in both
the time and transformed domains can lead to a new strategy of audio coding. Therefore,
besides fundamental research, project results should indicate whether the new techniques
could be potentially useful in practice in the future.
The project goal is to develop new models and techniques for audio restoration involving
perceptual information, which will lead to a better restoration quality in terms of objective and
subjective perceptual quality. Using the new unified framework, an experimental audio codec
will be proposed.
Keywords: Signal processing; audio; restoration; sparsity; inpainting; declipping; dequantization; coding;
auditory modeling; psychoacoustics; time-frequency representations; filterbanks
Signal Processing Laboratory, Faculty of Electrical Engineering and Communication, Brno University of Technology, Brno
doc. Mgr. Pavel Rajmic, Ph.D.
Ing. Pavel Záviška
Ing. Ondřej Mokrý
doc. Jiří Schimmel, Ph.D.
doc. RNDr. Vítězslav Veselý, CSc.