A Bayesian approach to voice activity detection using multiple statistical models and discriminative training

Tao Yu, John H. L. HansenInterspeech 2010

TL;DR

This paper proposes a more robust voice activity detection (VAD) system by using a Bayesian approach that combines multiple statistical models for speech and noise. Crucially, it introduces a discriminative training method to learn the optimal "mix" of these models for different noise environments, significantly improving VAD performance.

What Problem Does This Paper Address?

Voice Activity Detection (VAD) is a fundamental task in speech processing, aiming to distinguish between speech and silence (or noise). Accurate VAD is crucial for improving the performance of downstream applications like Automatic Speech Recognition (ASR) and speaker identification. Traditional VAD systems often rely on a single statistical model to represent speech and noise. However, real-world acoustic environments are highly diverse, meaning a single model may not accurately capture the characteristics of speech or noise across all conditions. This leads to performance degradation, particularly in challenging noisy environments. The paper addresses this by proposing a VAD system that can adapt to different noise conditions by intelligently combining multiple statistical models.

Methods

The core of the proposed method is a Bayesian hypothesis testing framework. Instead of assuming a single statistical model for speech (H1) and noise (H0), the authors consider a set of M candidate statistical models. The VAD decision is then made by combining the likelihoods from these models, weighted by their a priori probabilities.

The novelty lies in how these a priori probabilities are determined. The paper introduces a discriminative training approach based on the Minimum Classification Error (MCE) criterion. This method directly optimizes the VAD's performance by minimizing misclassification errors (miss-hits and false alarms) on a training dataset. The MCE criterion is approximated using a sigmoid function to allow for gradient-based optimization. The training process learns the optimal a priori probabilities for each candidate model under both the speech-active (H1) and speech-pause (H0) hypotheses, making the VAD system adaptive to specific noise environments.

The paper considers three candidate statistical models: Gaussian, Laplacian, and Gamma distributions, which have been shown to be effective for modeling speech and noise characteristics.

Datasets

The paper uses the AURORA database for evaluation. This database contains speech corrupted by six different noise conditions (babble, airport, car, station, exhibition, and restaurant) at approximately 5dB SNR. For each noise condition, the database provides 10 minutes of training data and 10 minutes of test data. The percentage of hand-marked speech-active frames in the dataset is 53.4%.

Key Insights and Ideas

  • Model Diversity is Key: A single statistical model is insufficient for robust VAD across diverse noise conditions. Combining multiple models offers a more flexible and accurate approach.
  • Adaptive Priors Matter: The a priori probabilities assigned to each statistical model are critical for VAD performance. These probabilities should be environment-dependent.
  • Discriminative Training for Optimal Mixing: MCE-based discriminative training is an effective way to learn these environment-dependent a priori probabilities, directly optimizing VAD performance by minimizing classification errors.
  • Noise-Specific Model Preferences: The learned optimal weights reveal interesting insights into which models are preferred under different hypotheses and noise conditions. For instance, the Laplacian model is often favored for speech-active segments, while the Gaussian model is preferred for speech-pause segments.

Key Results

The proposed Bayesian VAD with discriminative training significantly outperforms conventional VAD systems that use a single statistical model across all tested AURORA noise conditions. The paper presents Receiver Operating Characteristic (ROC) curves showing substantial improvements in Speech Hit Rate for a given False Alarm Rate. For example, the proposed method achieves better performance compared to single-model VADs (Gaussian, Laplacian, or Gamma) regardless of the initial model weights used for training. The stacked bar charts in Figure 2 illustrate how the optimal a priori probabilities vary across different noise conditions, demonstrating the adaptive nature of the proposed system.

Limitations and Future Work

  • Local Optimality: The authors acknowledge that the MCE training process can converge to local optima, especially when the initial weights are heavily biased towards a specific model (e.g., the Laplacian model in some noise conditions).
  • Parameter Transformation: To ensure non-negativeness and maintain constraints on the a priori probabilities during optimization, a parameter transformation is employed. The paper suggests this guarantees constraints are satisfied throughout iterations.
  • Future Work: Not explicitly detailed in the provided text, but the success of this approach suggests further exploration into more complex statistical models or more sophisticated discriminative training techniques could yield even greater improvements. The adaptive nature of the learned priors also opens avenues for real-time adaptation in dynamic acoustic environments.

Summary updated

BibTeX
@inproceedings{yu2010bayesian,
  author = {Tao Yu and John H. L. Hansen},
  title = {{A Bayesian approach to voice activity detection using multiple statistical models and discriminative training}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure