Temporal coherence and spectral regularity are critical cues for human auditory streaming processes and considered in many sound separation models. Some examples include the Conv-tasnet model, which focuses on temporal coherence using short length kernels to analyze sound, and the dual path convolution recurrent network (DPCRN) model, which uses two RNNs to analyze general patterns along the temporal and spectral dimensions on a spectrogram. In this paper, we propose a harmonic-aware tri-path convolution recurrent network (HATPCRN) model to separate singing voices on a spectrogram via the addition of an inter-band RNN, designed specifically for the harmonic structure. Evaluation results on the DSD100 and MUSDB18 datasets show that this addition can further boost separation performances of the DPCRN.

In this paper, we propose an ensemble learning based model to synthesize the logarithmic magnitude response of head-related transfer function (HRTF) using anthropometric features. We first cluster subjects based on relevant anthropometric features to reduce differences within each group, then we use the ensemble learning algorithm on clustered results to predict the log-magnitude HRTF. In the training phase, three deep neural networks (DNNs), each of which aims to predict log-magnitude HRTFs in a particular group, are trained using anthropometric and angle-related features. Afterward, another DNN is trained to integrate estimates from the three group-wise DNNs into log-magnitude HTRFs. The proposed model is compared with a baseline DNN model and our previously proposed model, which incorporates an auto-encoder for dimensionality reduction. Experimental results show that the proposed model performs the best in synthesizing log-magnitude HRTFs in terms of the log-spectral distortion (LSD) measure with great stability.