In conventional codebook-driven speech enhancement, only spectral envelopes of speech and noise are considered, and at the same time, the type of noise is the priori information when we enhance the noisy speech. In this paper, we propose a novel codebook-based speech enhancement method which exploits a priori information about binaural cues, including clean cue and pre-enhanced cue, stored in the trained codebook. This method includes two main parts: offline training of cues and online enhancement by means of cues. That is, we use the trained codebook to model a priori information of speech and noise offline and extract the pre-enhanced cue from the noisy observation online. The clean cue is estimated by the mapping of the weighted code vectors online, and the enhanced speech is produced by the estimated clean cue. The experimental results show that the proposed approach performs better than the reference methods in both stationary and non-stationary noise condition.
This article focus on the integral sliding mode control for a class of partially unknown T-S fuzzy systems on the basis of reinforcement learning technique. In the equivalent control, the control term is described as a form of feedback control based on the percompensation technique, such that the optimal control performance can be satisfied. Then the control gain is design by solving the algebraic Riccati equation according to the integral reinforcement learning (RL) method, and the requirements of ex-act information of system matrices is relaxed during the solving process. Base on the Lyapunov stability theory, the convergence and stability of the designed sliding mode control optimization scheme are analyzed. Finally, a numerical example is given to verified the effectiveness of the proposed theorems.
In this paper, an adaptive controller is investigated for a class of strict feedback nonlinear state constrained systems with both input delay and unknown control gains. To handle the unknown control gains and prevent constraint violation, we employ the integral Barrier Lyapunov Functions (iBLFs), which reduce some of the conservatism of transforming original state constraints into transformed error constraints. Pade approximation and an intermediate variable are applied to compensate the effect of the input delay. The unknown functions are approximated by Neural networks (NNs). It is proved that tracking error can converge to a compact set of the origin without violating the state constraint, and all closed loop signals remain bounded. The proposed controller can be verified by a numerical example.