In recent years, many researches focus on sound source localization based on neural networks, which is an appealing but difficult problem. In this paper, a novel time-domain end-to-end method for sound source localization is proposed, where the model is trained by two strategies with both cross entropy loss and mean square error loss. Based on the idea of multi-task learning, CNN is used as the shared hidden layers to extract features and DNN is used as the output layers for each task. Compared with SRP-PHAT, MUSIC and a DNN-based method, the proposed method has better performance.

The traditional eigen beam based localization algorithms are usually not employed on the non-spherical microphone array, for which the eigen beam is hard to be obtained. In this paper, the transfer functions are introduced to calculated the eigen beam on the non-spherical microphone array. Based on it, three localization algorithms including the eigen beam based intensity vector, eigen beam based beamforming, eigen beam based MUSIC, are employed and their performance on localization are evaluated.

To reduce the burden of storing and transmitting audio signals, they are often compressed with a lossy single-channel code. Because the high-frequency components are effectively truncated when using a low bitrate encoder, listeners may experience the sound as being uncomfortable, muffled, or dull. To compensate for the perceived degradation, bandwidth extension technology can be used to regenerate the missing high frequencies from the low-frequency components during the decoding process. In this paper the authors propose a bandwidth extension method based on Generative Adversarial Networks (GAN), which is used to estimate the relationship between the MDCT spectrum in the high-frequency part and the low-frequency part. It is evaluated by a discriminant network in the GAN to get a more natural result. A complete audio coding system was built by using AAC Low Complex as the single-channel core encoder with the proposed bandwidth extension method. To evaluate the audio quality decoded by the new system, a subjective evaluation experiment was carried out using the HE-AAC as the baseline system with the MUSHRA experimental method.

In this paper, a method for modeling distance dependent head-related transfer functions is presented. The HRTFs are first decomposed by spatial principal component analysis. Using deep neural networks, we model the spatial principal component weights of different distances. Then we realize the prediction of HRTFs in arbitrary spatial distances. The objective and subjective experiments are conducted to evaluate the proposed distance model and the distance variation function model, and the results have shown that the proposed model has less spectral distortions than distance variation function model, and the virtual sound generated by the proposed model has better performance in terms of distance localization.

The Ambisonic technique has been widely used for soundfield recording and reproduction recently. However, the basicAmbisonic decoding method will break down when the play-back loudspeakers distribute unevenly. Various methods havebeen proposed to solve this problem. This paper introducesseveral improvements to a recently proposed Ambisonic de-coding method, the matching projection method, for unevenloudspeaker layouts. The first improvement is energy preserv-ing; the second is introducing the “in-phase” weight, and thethird is introducing partial projection coefficients. To eval-uate the improved method, we compared it with the origi-nal one and the all-round Ambisonic decoding method witha 2-dimension unevenly arranged loudspeaker array. The re-sult shows our method greatly improves the original methodwhere the loudspeaker arranges very sparsely or densely.

The traditional multiple audio objects codec extracts the parameters of each object in the frequency domain and produces serious confusion because of high coincidence degree in subband among objects. This paper uses sparse domain instead of frequency domain and reconstruct audio object using the binary mask from the down-mixed signal based on the sparsity of each audio object. In order to overcome high coincidence degree of subband among different audio objects, the sparse autoencoder neural network is established. On this basis, a multiple audio objects codec system is built up. To evaluate this proposed system, the objective and subjective evaluation are carried on and the results show that the proposed system has the better performance than SAOC.