Abstract

Speech recognition and keyword detection are becoming increasingly popular applications for mobile systems. These applications have large memory and compute resource requirements, making their implementation on a mobile device quite challenging. In this paper, we design low cost neural network architectures for keyword detection and speech recognition. Wepresent techniques to reduce memory requirement by scaling down the precision of weight and biases without compromising on the detection/recognition performance. Experiments conducted on the Resource Management (RM) database show that for the keyword detection neural network, representing the weights by 5 bits results in a 6 fold reduction in memory compared to a floating point implementation with very little loss in performance. Similarly, for the speech recognition neural network, representing the weights by 6 bits results in a 5 fold reduction in memory while maintaining an error rate similar to a floating point implementation. Preliminary results in 40nm TSMC technology show that the networks have fairly small power consumption: 11.12mW for the keyword detection network and 51.96mW for the speech recognition network, making these designs suitable for mobile devices.

Original languageEnglish (US)
Pages (from-to)1-15
Number of pages15
JournalJournal of Signal Processing Systems
DOIs
StateAccepted/In press - Nov 25 2016

Fingerprint

Network Architecture
Network architecture
Speech recognition
Computer hardware
Speech Recognition
Fixed point
Hardware
Neural Networks
Neural networks
Data storage equipment
Resources
Mobile devices
Floating point
Mobile Devices
Fold
Requirements
Mobile Systems
Resource Management
Electric power utilization
Power Consumption

Keywords

  • Deep neural networks
  • Fixed-point architecture
  • Keyword detection
  • Memory compression
  • Speech recognition

ASJC Scopus subject areas

  • Control and Systems Engineering
  • Theoretical Computer Science
  • Signal Processing
  • Information Systems
  • Modeling and Simulation
  • Hardware and Architecture

Cite this

A Fixed-Point Neural Network Architecture for Speech Applications on Resource Constrained Hardware. / Shah, Mohit; Arunachalam, Sairam; Wang, Jingcheng; Blaauw, David; Sylvester, Dennis; Kim, Hun Seok; Seo, Jae-sun; Chakrabarti, Chaitali.

In: Journal of Signal Processing Systems, 25.11.2016, p. 1-15.

Research output: Contribution to journalArticle

Shah, Mohit ; Arunachalam, Sairam ; Wang, Jingcheng ; Blaauw, David ; Sylvester, Dennis ; Kim, Hun Seok ; Seo, Jae-sun ; Chakrabarti, Chaitali. / A Fixed-Point Neural Network Architecture for Speech Applications on Resource Constrained Hardware. In: Journal of Signal Processing Systems. 2016 ; pp. 1-15.
@article{057e64fb837a44e98bbbf16592a29345,
title = "A Fixed-Point Neural Network Architecture for Speech Applications on Resource Constrained Hardware",
abstract = "Speech recognition and keyword detection are becoming increasingly popular applications for mobile systems. These applications have large memory and compute resource requirements, making their implementation on a mobile device quite challenging. In this paper, we design low cost neural network architectures for keyword detection and speech recognition. Wepresent techniques to reduce memory requirement by scaling down the precision of weight and biases without compromising on the detection/recognition performance. Experiments conducted on the Resource Management (RM) database show that for the keyword detection neural network, representing the weights by 5 bits results in a 6 fold reduction in memory compared to a floating point implementation with very little loss in performance. Similarly, for the speech recognition neural network, representing the weights by 6 bits results in a 5 fold reduction in memory while maintaining an error rate similar to a floating point implementation. Preliminary results in 40nm TSMC technology show that the networks have fairly small power consumption: 11.12mW for the keyword detection network and 51.96mW for the speech recognition network, making these designs suitable for mobile devices.",
keywords = "Deep neural networks, Fixed-point architecture, Keyword detection, Memory compression, Speech recognition",
author = "Mohit Shah and Sairam Arunachalam and Jingcheng Wang and David Blaauw and Dennis Sylvester and Kim, {Hun Seok} and Jae-sun Seo and Chaitali Chakrabarti",
year = "2016",
month = "11",
day = "25",
doi = "10.1007/s11265-016-1202-x",
language = "English (US)",
pages = "1--15",
journal = "Journal of Signal Processing Systems",
issn = "1939-8018",
publisher = "Springer New York",

}

TY - JOUR

T1 - A Fixed-Point Neural Network Architecture for Speech Applications on Resource Constrained Hardware

AU - Shah, Mohit

AU - Arunachalam, Sairam

AU - Wang, Jingcheng

AU - Blaauw, David

AU - Sylvester, Dennis

AU - Kim, Hun Seok

AU - Seo, Jae-sun

AU - Chakrabarti, Chaitali

PY - 2016/11/25

Y1 - 2016/11/25

N2 - Speech recognition and keyword detection are becoming increasingly popular applications for mobile systems. These applications have large memory and compute resource requirements, making their implementation on a mobile device quite challenging. In this paper, we design low cost neural network architectures for keyword detection and speech recognition. Wepresent techniques to reduce memory requirement by scaling down the precision of weight and biases without compromising on the detection/recognition performance. Experiments conducted on the Resource Management (RM) database show that for the keyword detection neural network, representing the weights by 5 bits results in a 6 fold reduction in memory compared to a floating point implementation with very little loss in performance. Similarly, for the speech recognition neural network, representing the weights by 6 bits results in a 5 fold reduction in memory while maintaining an error rate similar to a floating point implementation. Preliminary results in 40nm TSMC technology show that the networks have fairly small power consumption: 11.12mW for the keyword detection network and 51.96mW for the speech recognition network, making these designs suitable for mobile devices.

AB - Speech recognition and keyword detection are becoming increasingly popular applications for mobile systems. These applications have large memory and compute resource requirements, making their implementation on a mobile device quite challenging. In this paper, we design low cost neural network architectures for keyword detection and speech recognition. Wepresent techniques to reduce memory requirement by scaling down the precision of weight and biases without compromising on the detection/recognition performance. Experiments conducted on the Resource Management (RM) database show that for the keyword detection neural network, representing the weights by 5 bits results in a 6 fold reduction in memory compared to a floating point implementation with very little loss in performance. Similarly, for the speech recognition neural network, representing the weights by 6 bits results in a 5 fold reduction in memory while maintaining an error rate similar to a floating point implementation. Preliminary results in 40nm TSMC technology show that the networks have fairly small power consumption: 11.12mW for the keyword detection network and 51.96mW for the speech recognition network, making these designs suitable for mobile devices.

KW - Deep neural networks

KW - Fixed-point architecture

KW - Keyword detection

KW - Memory compression

KW - Speech recognition

UR - http://www.scopus.com/inward/record.url?scp=84997079626&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84997079626&partnerID=8YFLogxK

U2 - 10.1007/s11265-016-1202-x

DO - 10.1007/s11265-016-1202-x

M3 - Article

SP - 1

EP - 15

JO - Journal of Signal Processing Systems

JF - Journal of Signal Processing Systems

SN - 1939-8018

ER -