Classification of Binaural Audio for Sound Immersion Assessment in TV 3.0

In recent years, artificial intelligence has advanced exponentially, allowing new approaches for evaluating sound immersion. This study explored a machine learning-based method to assess audio immersion by classifying binaural recordings captured using a dummy head. This classification aims to identify the type of reproduction system used, indirectly indicating the immersion level. A dataset was created by convolving test audio signals provided by the SBTVD Forum with binaural room impulse responses from the SADIE II database. The convolution process followed MPEG-H standards, simulating 2.0, 5.1, and 5.1 + 4H sound configurations required by the Brazilian TV 3.0 standard. The resulting binaural signals were segmented into 5-second samples and converted into image representations using spectrograms and phase-difference analyses. A convolutional neural network based on the VGG-16 architecture was trained to classify four audio formats: monophonic, stereophonic, surround 5.1, and 5.1 + 4H. Using transfer learning, the model achieved 92.21% accuracy for the validation data. These promising results demonstrate the feasibility of artificial intelligence-driven sound immersion evaluation, which offers a novel approach for immersive audio system classification.