基于ADTDNN的低资源语音识别方法研究  

Research on Low-Resource Speech Recognition Based on ADTDNN

作　　者：顾龙昊 黄连丽[1] 周奎 张子越 GU Longhao;HUANG Lianli;ZHOU Kui;ZHANG Ziyue(Department of Electrical and Information Engineering,Hubei University of Automotive Technology;Sharing-X Key Joint Laboratory,Institute of Automotive Engineers,Hubei University of Automotive Technology,Shiyan 442002,China)

机构地区：[1]湖北汽车工业学院电气与信息工程学院 [2]湖北汽车工业学院汽车工程师学院Sharing-X重点联合实验室,湖北十堰442002

出　　处：《软件导刊》2024年第9期76-81,共6页Software Guide

基　　金：湖北省重点研发计划项目(2023BAB169);湖北省武汉市科技重大专项(2022013702025184)。

摘　　要：为解决低资源条件下由于训练数据不足导致识别精度降低、泛化能力较差的问题,提出一种语音识别方法。该方法利用卷积池化提取特征信息,将Attention机制与DTDNN融合成为ADTDNN,以提升低资源环境下模型捕捉序列中关键信息的能力;采用链接时序分类简化模型的识别流程;使用Transformer作为语言模型。在Aishell-1数据集上的实验结果表明,低资源环境下基于ADTDNN的语音识别模型与LAS、Transformer等主流端到端模型相比,字错误率分别降低了3.7%和1.0%。A speech recognition approach has been proposed to address the problem of reduced recognition accuracy and poorer generalization performance due to insufficient training data in low-resource conditions.This method leverages convolutional neural networks to extract feature information.It combines the attention mechanism with delayed time-delay neural networks,referred to as ADTDNN,enhancing the model′s ability to capture key information in sequences within low-resource environments.The approach employs linking temporal classification to streamline the recognition process of the model.Additionally,a Transformer is utilized as the language model.Experimental results on the Aishell-1 dataset demonstrate that the ADTDNN-based speech recognition model in low-resource settings reduces word error rates by 3.7%and 1%compared to mainstream end-to-end models like LAS and Transformer,respectively.

关 键 词：语音识别 时延神经网络 TRANSFORMER 数据增强 低资源 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]