Estimating noise from noisy speech features with a monte carlo variant of the expectation maximization algorithm

Friedrich Faubel, Dietrich KlakowInterspeech 2010 · DOI

Abstract

Robust speech recognition in noisy environments requires accurate estimation of background noise parameters. Traditional Expectation-Maximization (EM) methods often rely on linear approximations like Vector Taylor Series, which can be computationally expensive and inaccurate. This paper introduces a Monte Carlo EM (MC-EM) approach that utilizes importance sampling and Parzen-window density estimation to numerically approximate noise distributions. Experiments on simulated and real-world car noise demonstrate that MC-EM achieves up to a 34% relative reduction in word error rate while operating significantly faster than conventional Gaussian Mixture EM methods.

Introduction

Automatic Speech Recognition (ASR) systems frequently encounter performance degradation in real-world "wild" environments, such as voice assistants operating inside moving vehicles. To maintain accuracy, these systems must accurately model and subtract background noise. Traditionally, the Expectation-Maximization (EM) algorithm has been the standard for estimating noise parameters. However, the "E-step" involves solving complex, non-linear integrals that describe the interaction between speech and noise. Because these integrals lack closed-form solutions, researchers have historically relied on approximations like Vector Taylor Series (VTS) or Unscented Transforms. These methods often fail to capture the true noise distribution, leading to slow convergence and suboptimal recognition accuracy.

The Monte Carlo EM Approach

To overcome the limitations of analytical approximations, this research proposes a Monte Carlo EM (MC-EM) framework. Instead of simplifying the mathematical interaction, the algorithm approximates the required integrals numerically through sampling.

The process involves three core components:

  1. Importance Sampling: The algorithm draws random noise samples and weights them based on their likelihood of producing the observed noisy speech, given a pre-trained clean speech model.
  2. Parzen-Window Density Estimation: This technique allows the system to simulate the distribution of noisy speech under specific noise conditions without requiring a closed-form solution, effectively breaking the Gaussian constraint of previous methods.
  3. Iterative Refinement: The process is iterative. Each step uses the updated noise estimate to improve the quality of the samples in the next iteration, leading to a highly accurate environmental noise model.

Experimental Evaluation

The MC-EM method was evaluated using two datasets:

  • Simulated Noise: The MC-WSJ-AV (Wall Street Journal) corpus was augmented with "destroyer engine" and "factory" noises from the NOISEX-92 database at various signal-to-noise ratios (SNRs).
  • Real-World Noise: The AVICAR corpus provided recordings of phone numbers in actual driving conditions (up to 55 mph) with various window configurations.

Key Results and Insights

The MC-EM approach demonstrated superior performance in both accuracy and computational efficiency. Unlike traditional Gaussian Mixture EM (GM-EM) methods that force non-linear transformations into Gaussian shapes, MC-EM transforms the entire distribution via sampling, preserving the true data geometry.

Word Error Rate (WER) vs. iteration for simulated engine and factory noise at different SNR levels. Solid lines represent MC-EM, while dashed lines represent the GM-EM baseline.

As shown in the simulated noise results, MC-EM consistently outperformed the baseline across all SNR levels. Specifically, it achieved a 34% relative reduction in WER for engine noise and a 16% reduction for factory noise. Furthermore, the Monte Carlo approach was found to be up to 10 times faster than GM-EM using full covariance matrices, as it avoids the heavy cost of propagating individual Gaussian components through non-linear functions.

In real-world tests using the AVICAR "windows down" dataset, the MC-EM method showed significant robustness and convergence.

WER comparison between MC-EM and GM-EM over iterations for the AVICAR real-world car noise dataset.

The baseline WER in this challenging environment was 15.54%. As seen in the table above, MC-EM successfully reduced this to 10.48% by the sixth iteration, while the GM-EM baseline struggled to improve significantly from its initial state.

Conclusion and Future Work

The MC-EM algorithm provides a robust and efficient alternative for noise estimation in ASR. By replacing rigid mathematical approximations with flexible numerical sampling, it achieves better accuracy and faster execution. While the method currently depends on a pre-existing clean speech distribution, future work could explore integrating these sampling techniques into modern deep-learning-based speech enhancement front-ends to further improve performance across diverse speakers and accents.

Summary updated

BibTeX
@inproceedings{faubel2010estimating,
  author = {Friedrich Faubel and Dietrich Klakow},
  title = {{Estimating noise from noisy speech features with a monte carlo variant of the expectation maximization algorithm}},
  booktitle = {Interspeech},
  year = {2010},
  doi = {10.21437/interspeech.2010-578},
}
Figure