|COMANDUCCI LUCA||Cycle: XXXIV |
Tutor: CESANA MATTEO
Advisor: SARTI AUGUSTO Major Research topic
:SOUNDfield reconstruction from diStributed microPHones using dEep leaRning tEchniqueS (SOUND SPHERES) Abstract:
Space-time audio processing is a research field that focuses on a wide range of applications and allows to analyze the soundfield in an environment, usually through the adoption of setups based on microphone arrays.
Classical space-time audio processing techniques rely on complex models of the considered environments and on strong assumptions related to the property of the audio signals and noise. These assumptions strongly condition the application of such techniques to settings similar to the ones used to build the model. Also they usually imply the need of very large setups.
In the last years there has been a growing interest towards the combination of space-time audio techniques with machine learning, since it makes possible to overcome the aforementioned limitations. In particular it allows us to avoid an explicit analysis of the environment and instead to derive it in a data-driven fashion, directly from the input audio signals.
It is possible to envision a framework able to reconstruct the soundfield using minimal setups, such as ones consisting of distributed single microphones. This will enable us to solve a series of space-time audio processing-related problems, with minimal setups. This objective is challenging since by using single distributed microphones, we are not able to properly sample the acoustic field and we incur in problems related to spatial aliasing. In order to recover the missing information it is possible to apply deep learning techniques to the acquired audio signals.
Since the soundfield is usually acquired by means of microphone arrays, we need to choose a representation for the array data best suited at working with machine learning. The Ray Space Transform (RST) extends the plenacoustic function to the directional case and parametrizes the soundfield in the domain of the acoustic rays, i.e. oriented lines along which the acoustic radiance is constant. The great advantage of the RST representation is that acoustic objects of interest are represented by linear patterns, thus space-time processing problems can be treated with pattern analysis techniques. This linear representation of the soundfield also makes the RST particularly suited at being analyzed through deep learning techniques.