面向低功耗语音识别的可重构数据通路架构设计与实现

 2022-05-16 20:33:41

论文总字数:30635字

摘 要

计算能效一直是制约高精度低功耗语音识别发展的瓶颈之一,特别是当神经网络技术运用于智能语音识别设备时,计算架构的能效问题成为首要考虑的关键点,因此在识别精度满足设计需求的条件下,更低功耗的智能语音识别架构和电路设计具有重要意义。基于关键词语音识别(KWS)系统,本文设计并实现了一种面向低功耗语音识别的可重构数据通路架构。

论文主要内容包括:(1)针对基于卷积神经网络的关键词识别系统,探究了不同噪声类型以及不同信噪比场景下,基于神经网络的分类模块对输入数据和权重数据位宽的需求特征,分析了不同数据位宽对语音识别系统总体识别率的影响关系。(2)实现了可重构数据通路的设计,其可以动态地重新配置计算单元的数据位宽以适应不同的计算精度要求,与标准计算单元相比,可显著降低能耗。

本文基于TSMC 28nm工艺完成了面向关键词识别电路设计。仿真测试结果表明:关键词识别系统的功耗最低可达到0.6mW,其峰值能效达到34TOPS/W,在不同的噪声背景下,系统识别率可达到94.6%。

关键词: 关键词识别,卷积神经网络,自适应位宽,低功耗

Abstract

Computational efficiency has always been one of the problems restricting the development of high precision and low power speech recognition. Especially when neural network technology is applied to intelligent speech recognition devices, the energy efficiency of computing architecture becomes the key point. Therefore, under the condition that the recognition accuracy meets the design requirements, the intelligent speech recognition architecture and circuit design with lower power consumption are of great significance. Based on key word speech recognition (KWS) system, this paper designs and realizes a reconfigurable data path architecture for low-power speech recognition.

The main work of the thesis include: (1) Aiming at the key word recognition system based on convolutional neural network, the demand characteristics of the classification module based on neural network for input data and weighted data bit widths under different noise types and different SNR scenarios are explored. Besides, the influence of different data bit widths on the overall recognition rate of speech recognition system is analyzed. (2) The design of reconfigurable data path is realized, which can dynamically reconfigure the data bit width of the computing cell to meet different requirements of calculation accuracy. This work can significantly reduce energy consumption compared with the standard computing cell.

Based on TSMC 28nm technology, the circuit design for keyword recognition is completed. The simulation results show that the power consumption can reach 0.6mW at the TT process angle and 25°C, and the peak energy efficiency reaches 34TOPS/W. Under different noise backgrounds, the recognition rate can reach 94.6%.

KEY WORDS:keyword recognition, Convolutional Neural Network, adaptive bit width, low power consumption

目 录

摘 要 I

Abstract II

第一章 绪论 1

1.1研究背景与意义 1

1.2国内外研究现状 2

1.2.1深度卷积神经网络发展历程及研究现状 2

1.2.2神经网络位宽调节研究现状 3

1.3本文研究内容及创新点 3

1.3.1本文研究内容 3

1.3.2本文创新点 4

1.4论文组织结构 4

第二章 面向关键词识别的卷积神经网络算法分析 5

2.1关键词识别系统基本框架 5

2.2语音特征提取算法分析 6

2.2.1典型特征提取算法分析比较 6

2.2.2 MFCC特征提取算法分析 6

2.3基于卷积神经网络的语音识别方法分析 8

2.3.1卷积神经网络算法分析 9

2.3.2卷积神经网络数据重用分析 12

2.4本章小结 13

第三章 权重及数据位宽量化方案分析设计 14

3.1语音识别系统神经网络结构分析设计 14

3.2卷积神经网络正向以及反向量化策略 15

3.2.1基于正向传播和反向梯度更新的权重量化 15

3.2.2基于正向传播和反向梯度更新的数据量化 17

3.3量化位宽分析 17

3.4本章小结 20

第四章 可重构数据通路架构设计 21

4.1关键词识别系统整体架构设计 21

4.2卷积网络整体层次访存模型设计 21

4.3卷积计算数据通路设计 23

4.3.1访存控制单元设计 23

4.3.2计算阵列数据通路设计 24

4.4计算单元设计 25

4.5计算架构验证与结果分析 26

4.5.1寄存器传输级(RTL)仿真验证 26

4.5.2性能测试与方案对比 27

4.6本章小结 28

第五章 总结与展望 29

5.1总结 29

5.2展望 29

致 谢 30

参考文献 31

绪论

剩余内容已隐藏,请支付后下载全文,论文总字数:30635字

您需要先支付 80元 才能查看全部内容!立即支付

该课题毕业论文、开题报告、外文翻译、程序设计、图纸设计等资料可联系客服协助查找;