面向低功耗语音识别的近似乘法器设计与实现

论文总字数：30323字

摘 要

本文研究了基于精度可控近似乘法器的高能效计算方案，大幅降低电路功耗，能够在适配不同应用场景时，实现精度受限范围内的高能效计算。

首先分析了KWS实现的各种方案，并对各种语音提取方案进行分析研究，通过对比确定了神经网络与MFCC特征提取的方案。

接着，基于近似计算的专用设计方法，利用电路与系统的容错性，使用非精确的硬件设计实现高效能。通过分析迭代对数、量化分解等各类近似乘法运算的原理，考虑各类近似乘法方案的计算精度变化曲线特征，分析了各种实现近似计算架构的优点和缺点，并评估其计算性能与功耗开销。通过对近似计算电路结构以及精度控制方法进行协同仿真、方案评估优化，实现满足不同噪声类型、不同SNR场景下低功耗语音识别近似乘法器电路和精度控制方法设计。设计了基于近似加法的近似乘法器，并研究了相应的误差修复电路。

最后研究了CNN加速器架构，基于SNR的精度控制以及乘法器流程，并在语音库中对设计的方案进行了相应测试训练。证明了设计的正确合理性。

关键词：语音识别，近似乘法器，低功耗，高能效

Abstract

In this paper, an energy-efficient calculation scheme based on precision controllable approximate multiplier is studied, which greatly reduces the power consumption of the circuit, and can achieve high-energy efficiency calculation within the limited range of accuracy when adapting to different application scenarios.

Firstly, the various schemes of KWS implementation are analyzed, and various speech extraction schemes are analyzed and studied. The schemes of neural network and MFCC feature extraction are determined by comparison.

Then, based on the special design method of approximate calculation, using the fault tolerance of the circuit and the system, using the inaccurate hardware design to achieve high performance. By analyzing the principles of various approximation multiplication operations such as iterative logarithm and quantization decomposition, considering the characteristics of the calculation accuracy curve of various approximate multiplication schemes, the advantages and disadvantages of various approximate calculation architectures are analyzed, and the computational performance and work are evaluated. Cost. Through the co-simulation and scheme evaluation of the approximate calculation circuit structure and the precision control method, the design of the approximate multiplier circuit and the precision control method for low-power speech recognition under different noise types and different SNR scenarios is realized. An approximate multiplier based on approximation addition is designed and the corresponding error repair circuit is studied.

Finally, the CNN accelerator architecture, SNR-based precision control and multiplier flow are studied, and the design scheme is tested and trained in the speech library. Prove the correctness and rationality of the design.

KEY WORDS: Speech recognition, approximate multiplier, low power, high energy efficiency

目录

第一章 绪论 1

1.1研究背景 1

1.1.1深度学习的基本概念及其发展 1

1.1.2人工神经网络 1

1.2国内外研究现状 2

1.3论文研究内容及创新点 3

1.3.1研究内容 3

1.3.2创新点 3

1.4论文组织结构 3

第二章 面向语音识别的神经网络算法分析 4

2.1关键词识别典型算法分析及对比 4

2.2语音识别的特征提取算法 4

2.2.1 主流的特征提取算法 5

2.2.2 基于 MFCC 的特征提取算法分析 8

2.3 卷积、递归、全连接神经网络结构特征分析 12

2.4 硬件结构 15

2.5 本章小结 17

第三章 近似乘法计算分析及比较 17

3.1 神经网络中近似计算 17

3.2典型乘法计算及其结构 18

3.2.1对数乘法计算 18

3.2.2 截断乘法计算 19

3.2.3 共享乘法计算 19

3.3 基于近似加法的近似乘法计算 20

第四章 基于CNN网络系统实现 24

4.1 面向KWS的CNN 加速器架构设计 24

4.2 关键词识别 CNN 实现 24

4.3乘法器工作流程 29

第五章 总结与展望 33

5.1总结 33

5.2展望 33

致 谢 34

参考文献 35

第一章 绪论

1.1研究背景

1.1.1深度学习的基本概念及其发展

剩余内容已隐藏，请支付后下载全文，论文总字数：30323字