Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Resource-Efficient Speech Quality Prediction through Quantization Aware Training and Binary Activation Maps

Mattias Nilsson, Riccardo Miccini, Clément Laroche, Tobias Piechowiak, Friedemann Zenke

TL;DR

This work investigates binary activation maps for speech quality prediction on a convolutional architecture based on DNSMOS and shows a path toward substantial resource savings by supporting mixed-precision binary multiplication in hard- and software.

Abstract

As speech processing systems in mobile and edge devices become more commonplace, the demand for unintrusive speech quality monitoring increases. Deep learning methods provide high-quality estimates of objective and subjective speech quality metrics. However, their significant computational requirements are often prohibitive on resource-constrained devices. To address this issue, we investigated binary activation maps (BAMs) for speech quality prediction on a convolutional architecture based on DNSMOS. We show that the binary activation model with quantization aware training matches the predictive performance of the baseline model. It further allows using other compression techniques. Combined with 8-bit weight quantization, our approach results in a 25-fold memory reduction during inference, while replacing almost all dot products with summations. Our findings show a path toward substantial resource savings by supporting mixed-precision binary multiplication in hard- and software.

Resource-Efficient Speech Quality Prediction through Quantization Aware Training and Binary Activation Maps

TL;DR

This work investigates binary activation maps for speech quality prediction on a convolutional architecture based on DNSMOS and shows a path toward substantial resource savings by supporting mixed-precision binary multiplication in hard- and software.

Abstract

As speech processing systems in mobile and edge devices become more commonplace, the demand for unintrusive speech quality monitoring increases. Deep learning methods provide high-quality estimates of objective and subjective speech quality metrics. However, their significant computational requirements are often prohibitive on resource-constrained devices. To address this issue, we investigated binary activation maps (BAMs) for speech quality prediction on a convolutional architecture based on DNSMOS. We show that the binary activation model with quantization aware training matches the predictive performance of the baseline model. It further allows using other compression techniques. Combined with 8-bit weight quantization, our approach results in a 25-fold memory reduction during inference, while replacing almost all dot products with summations. Our findings show a path toward substantial resource savings by supporting mixed-precision binary multiplication in hard- and software.
Paper Structure (15 sections, 5 equations, 3 figures, 1 table)

This paper contains 15 sections, 5 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Problem Definition
  4. Quantization
  5. Binary Activation Maps (BAMs)
  6. Binary Weights
  7. 8-Bit Quantization
  8. Surrogate Gradient
  9. Experimental Setup
  10. Model
  11. Dataset
  12. Training and Evaluation Metrics
  13. Results
  14. Discussion
  15. Acknowledgements

Figures (3)

  • Figure 1: Surrogate derivative for training of neural networks with binary activations. The surrogate activation function is a fast sigmoid with adjustable steepness.
  • Figure 2: Evaluation metrics. Test and for the following different quantizations of DNSMOS: binary of activation maps (gray), with (orange), and 8-bit of activations and weights (green). The black lines indicate standard deviations from repeated instances of training (n=4).
  • Figure 3: Baseline model predictions vs. and vs. baseline model predictions on the test set.