Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dynamic Gated Recurrent Neural Network for Compute-efficient Speech Enhancement

Longbiao Cheng, Ashutosh Pandey, Buye Xu, Tobi Delbruck, Shih-Chii Liu

TL;DR

Test results obtained from several state-of-the-art compute-efficient RNN-based speech enhancement architectures using the DNS challenge dataset, show that the D-GRU based model variants maintain similar speech intelligibility and quality metrics comparable to the baseline GRU based models even with an average 50% reduction in GRU computes.

Abstract

This paper introduces a new Dynamic Gated Recurrent Neural Network (DG-RNN) for compute-efficient speech enhancement models running on resource-constrained hardware platforms. It leverages the slow evolution characteristic of RNN hidden states over steps, and updates only a selected set of neurons at each step by adding a newly proposed select gate to the RNN model. This select gate allows the computation cost of the conventional RNN to be reduced during network inference. As a realization of the DG-RNN, we further propose the Dynamic Gated Recurrent Unit (D-GRU) which does not require additional parameters. Test results obtained from several state-of-the-art compute-efficient RNN-based speech enhancement architectures using the DNS challenge dataset, show that the D-GRU based model variants maintain similar speech intelligibility and quality metrics comparable to the baseline GRU based models even with an average 50% reduction in GRU computes.

Dynamic Gated Recurrent Neural Network for Compute-efficient Speech Enhancement

TL;DR

Test results obtained from several state-of-the-art compute-efficient RNN-based speech enhancement architectures using the DNS challenge dataset, show that the D-GRU based model variants maintain similar speech intelligibility and quality metrics comparable to the baseline GRU based models even with an average 50% reduction in GRU computes.

Abstract

This paper introduces a new Dynamic Gated Recurrent Neural Network (DG-RNN) for compute-efficient speech enhancement models running on resource-constrained hardware platforms. It leverages the slow evolution characteristic of RNN hidden states over steps, and updates only a selected set of neurons at each step by adding a newly proposed select gate to the RNN model. This select gate allows the computation cost of the conventional RNN to be reduced during network inference. As a realization of the DG-RNN, we further propose the Dynamic Gated Recurrent Unit (D-GRU) which does not require additional parameters. Test results obtained from several state-of-the-art compute-efficient RNN-based speech enhancement architectures using the DNS challenge dataset, show that the D-GRU based model variants maintain similar speech intelligibility and quality metrics comparable to the baseline GRU based models even with an average 50% reduction in GRU computes.
Paper Structure (13 sections, 7 equations, 2 figures, 2 tables)

This paper contains 13 sections, 7 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Dynamic Gated RNN (DG-RNN)
  3. Dynamic GRU (D-GRU)
  4. gru grugru Update
  5. dgru dgrudgru Update
  6. Cost Savings
  7. Experimental Setup
  8. dgru dgrudgru Based Speech Enhancement
  9. Dataset
  10. Evaluation Metrics
  11. Results
  12. Related Works and Discussions
  13. Conclusion

Figures (2)

  • Figure 1: Illustration of the update processes of (A) conventional Recurrent Neural Network (RNN) and (B) Dynamic Gated Recurrent Neural Network (DG-RNN) at step $t$. (A) For conventional RNN, all neurons in the hidden state are updated at each step. (B) DG-RNN first identifies which neurons need updating, as indicated by a $1$ in the proposed select gate $\boldsymbol{g}_t$. When a neuron is marked with $1$, it undergoes the RNN update process. Those marked with $0$ retain their values from the previous hidden state.
  • Figure 2: Mean PESQ improvement ($\Delta_{PESQ}$) of different update percentages $\mathcal{P}$ on different models under {-5, 0, 5} dB SNRs.