End-to-end sound field reproduction based on deep learning

Area of Science:

• Acoustics and Signal Processing
• Artificial Intelligence in Audio Engineering
• Virtual Reality Technologies

Background:

• Sound field reproduction is crucial for immersive virtual reality experiences.
• Accurate reproduction requires precise calculation of loudspeaker driving signals based on microphone inputs and environmental factors.
• Conventional methods face challenges, especially in achieving high-frequency fidelity.

Purpose of the Study:

• To propose an end-to-end deep learning method for sound field reproduction.
• To enhance the accuracy and high-frequency performance of virtual acoustic environments.
• To compare the proposed method against established techniques.

Main Methods:

• Developed a deep learning model using a convolutional autoencoder with frequency-domain skip connections.
• Incorporated sparse layers to effectively capture sound field characteristics.
• Utilized sound pressure signals from microphones as input and loudspeaker driving signals as output.

Main Results:

• The proposed deep learning method significantly reduced reproduction errors compared to pressure matching and LASSO methods.
• Superior performance was observed particularly at high frequencies.
• Experimental validation confirmed enhanced high-frequency reproduction accuracy for both single and multiple primary source scenarios.

Conclusions:

• The end-to-end deep learning approach offers a promising advancement in sound field reproduction.
• The method excels in high-frequency accuracy, addressing a key limitation of conventional techniques.
• This technology has the potential to significantly improve the realism of virtual acoustic environments.