近年来,深度学习方法在自动语音识别(ASR)中取得了优于传统机器学习方法的成果,尤其是基于长短时记忆(LSTM)的深度学习方法改善了ASR性能,然而传统LSTM对于处理连续输入流存在限制,需要大量内存带宽和计算资源。本研究提出了一种增强型深度学习LSTM循环神经网络(RNN)模型,该模型将RNN作为“遗忘门”,融入记忆块,能够在子序列开始时重置单元状态,从而高效处理连续输入流。此外,修改了LSTM网络的标准架构,以更有效地利用模型参数。与基于CNN和顺序模型的方法相比,在电网调度语音数据集上,LSTM-RNN模型以99.36%的准确率优于其他深度学习模型。

In recent years,deep learning methods have achieved superior results compared to traditional machine learning methods in Automatic Speech Recognition (ASR).Particularly,deep learning methods based on Long Short-Term Memory (LSTM) have improved ASR performance.However,traditional LSTM models have limitations when it comes to handling continuous input streams,requiring substantial memory bandwidth and computational resources.This research introduces an enhanced deep learning model called LSTM Recurrent Neural Network (RNN),which addresses these limitations.In this model,the RNN serves as a “forgetting gate” integrated into the memory block.It allows for efficient processing of continuous input streams by resetting the cell state at the beginning of each subsequence.Additionally,modifications have been made to the standard architecture of the LSTM network to more effectively utilize model parameters.Compared to methods based on CNN and sequential models,the LSTM-RNN model achieves a 99.36% accuracy on the power grid dispatch speech dataset,surpassing other deep learning models.