Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

U-DREAM: Unsupervised Dereverberation guided by a Reverberation Model

Louis Bahrman, Marius Rodrigues, Mathieu Fontaine, Gaël Richard

Abstract

This paper explores the outcome of training state-of-the-art dereverberation models with supervision settings ranging from weakly-supervised to virtually unsupervised, relying solely on reverberant signals and an acoustic model for training. Most of the existing deep learning approaches typically require paired dry and reverberant data, which are difficult to obtain in practice. We develop instead a sequential learning strategy motivated by a maximum-likelihood formulation of the dereverberation problem, wherein acoustic parameters and dry signals are estimated from reverberant inputs using deep neural networks, guided by a reverberation matching loss. Our most data-efficient variant requires only 100 reverberation-parameter-labeled samples to outperform an unsupervised baseline, demonstrating the effectiveness and practicality of the proposed method in low-resource scenarios.

U-DREAM: Unsupervised Dereverberation guided by a Reverberation Model

Abstract

This paper explores the outcome of training state-of-the-art dereverberation models with supervision settings ranging from weakly-supervised to virtually unsupervised, relying solely on reverberant signals and an acoustic model for training. Most of the existing deep learning approaches typically require paired dry and reverberant data, which are difficult to obtain in practice. We develop instead a sequential learning strategy motivated by a maximum-likelihood formulation of the dereverberation problem, wherein acoustic parameters and dry signals are estimated from reverberant inputs using deep neural networks, guided by a reverberation matching loss. Our most data-efficient variant requires only 100 reverberation-parameter-labeled samples to outperform an unsupervised baseline, demonstrating the effectiveness and practicality of the proposed method in low-resource scenarios.

Paper Structure

This paper contains 37 sections, 31 equations, 8 figures.

Table of Contents

  1. Introduction
  2. Related works
  3. Reverberation model
  4. Room Impulse Response Model
  5. Convolution in the Time-Frequency Domain
  6. The U-DREAM model
  7. Problem formulation
  8. Overview of the method
  9. Reverberation sampler $\mathcal{R}$
  10. Acoustic Analyzer $\mathcal{A}$
  11. Non-blind analysis
  12. Blind Analysis
  13. Reverberation Matching Loss and variants
  14. Loss variants
  15. Balancing the Loss Terms
  16. ...and 22 more sections

Figures (8)

  • Figure 1: Overview of the proposed method
  • Figure 2: Dereverberation with strong supervision: Comparison of the proposed training loss (supervision by the RIR) and the baseline training loss (supervision by the dry signal). Results are presented as the relative improvement compared to the reverberant input. The 95 confidence intervals are indicated by black lines.
  • Figure 3: Weak dereverberation: Comparison of the weakly-supervised models with the WPE method. All models (BiLSTM, FSN, TFL) are trained with the "Single" reverberation matching loss variant. Results are presented as the relative improvement compared to the reverberant input. The 95 confidence intervals are indicated by black lines.
  • Figure 4: Weak dereverberation: Comparison of the RM loss variants on the EARS-Reverb dataset. Results are presented as the relative improvement compared to the reverberant input.
  • Figure 5: Weak dereverberation: Degradation caused by training using weak supervision compared to strong supervision on the EARS-Reverb dataset, using the "Single" loss variant.
  • ...and 3 more figures