Research and Application of Power Dispatching Speech Recognition Based on Deep Neural Networks

Due to the computational resource consumption of the decoding module and the impact of noisy environment on speech recognition system, this paper designs a real-time multi-stage speech preprocessing algorithm. The algorithm changes the traditional streaming recognition mode of detecting endpoints while recognizing. It reduces the decoding power consumption of the system by first performing endpoint detection and then decoding module. For wake-up decoding and recognition decoding, the real-time multilevel speech preprocessing algorithm significantly reduces the real-time rate and power consumption of the entire speech recognition system, and at the same time ensures that the impact on the wake-up rate and word error rate is negligible. The real-time rate is reduced by 27%-37% and the power consumption of the whole system is reduced by about 25%.