Search

Millimeter wave (mmWave) communications are considered as a key technology for 5G due to the huge spectrum resources which can't be provided in traditional frequency band. However, traditional full digital precoding methods are difficult to implement in massive multiple-input-multiple-output (MIMO) systems due to high cost and high power consumption. Nevertheless, mmWave hybrid precoding architecture achieves a good balance between performance and cost. Inspired by the matching pursuit (MP) algorithm, we propose a near-optimal hybrid precoding algorithm for fully-connected hybrid architecture. Instead of searching the analog vectors from the set of the array response vectors, we successively construct the appropriate analog precoding vectors by the alternating iteration method to constitute the analog precoding matrix, and the baseband precoding matrix is derived by regular least squares. In addition, the proposed algorithm takes no assumptions or approximations. Numerical simulations show that the proposed algorithm outperforms previous methods in performance with lower computational complexity.

In multi-user millimeter wave (mmWave) communications, massive multiple-input multiple-output (MIMO) systems can achieve high gain and spectral efficiency significantly. To reduce the hardware complexity and energy consumption of massive MIMO systems, hybrid precoding as a crucial technique has attracted extensive attention. Most previous works for hybrid precoding developed algorithms based on optimization or exhaustive search approaches that either lead to sub-optimal performance or have high computational complexity. Motivated by the thought of cross-fertilization between Data-driven and Model-driven approaches, we consider applying deep learning approach and introduce the Hybrid Precoding Network(HPNet), which is a compressed deep neural network exploiting the feature extracting (thanks to convolutional kernels) and generalization ability of neural networks and the natural sparsity of mmWave channels. The HPNet takes imperfect channel state information (CSI) as the input and predicts the analog precoder and baseband precoder for multi-user massive MIMO systems. Moreover, in order to make the approach more practical in real scenarios, we further introduce a model compression algorithm, using network pruning, to greatly reduce the computational complexity and memory usage of the neural network while almost retaining the model performance and then assess the influence of pruned parameters in the network. Numerical experiments demonstrate that HPNet outperforms state-of-the-art hybrid precoding schemes with higher performance and stronger robustness. Finally, we analyze and compare the computational complexity of different schemes.

Accurate estimation of reverberation time T_60 proved to have a positive effect on the automatic speech recognition (ASR) used in the voice-controlled devices and the reconstruction of the acoustic field. Recently, researchers have proposed some algorithms to estimate T_60. However, few of them directly use the spatial information about the acoustic environment contained in the speech for accurate T_60 estimation. We propose a deep learning approach as a regression problem to use binaural reverberant speech generated by the clean speech convolved with simulated Room Impulse Response (RIR) to estimate T_60. Adaptive cIRM estimator firstly estimates the complex Ideal Ratio Mask (cIRM), which is strongly correlated with T_60, and then a CNN-based T_60 estimator is used to estimate T_60 with cIRM. The experimental results show that our proposed approach outperforms the state-of-the-art method of T_60 estimation.

In this paper, we propose an end to end model for keyword spotting(KWS) with densely connected convolutional network(DenseNet) and the integration of Short Time Fourier Transform(STFT) which aims to extract utterances from raw waveform directly. Furthermore, we investigate the efficiency of adaptive and trainable window function in the task of keywords spotting. Using the recently-released Google Speech Commands Dataset as our benchmark. Our DenseNet implementation significantly outperforms previous neural networks based KWS system in terms of accuracy.