Improvements of the beta-order minimum mean-square error (MMSE) spectral amplitude estimator using chi priors

Marek Trawicki, Michael JohnsonInterspeech 2012

TL;DR

This paper introduces an improved speech enhancement algorithm by combining a flexible "beta-order" cost function with a more sophisticated "Chi" statistical model for speech priors. This new approach leads to significant gains in both speech clarity (SSNR) and perceptual quality (PESQ) compared to existing methods, particularly in noisy environments.

What Problem Does This Paper Address?

Speech enhancement algorithms aim to remove unwanted background noise from audio signals while preserving the original speech as much as possible. A key challenge in this field is accurately estimating the spectral amplitude of the clean speech signal from noisy observations. Traditional methods often rely on simplified statistical models (like the Rayleigh distribution) for speech priors, which can lead to trade-offs between noise reduction and speech distortion, such as the introduction of "musical tones" (unnatural, repeating artifacts). While some advancements have introduced more flexible cost functions (like the beta-order MMSE) or more advanced statistical models (like the Chi prior), these improvements haven't been combined into a single, unified estimator. This paper addresses this gap by merging the benefits of the beta-order cost function with the Chi speech prior to achieve better overall speech enhancement.

Methods

The core of this paper is the derivation of a new MMSE Beta-Order Spectral Amplitude estimator that utilizes Chi statistical models for the speech prior.

  • Beta-Order Cost Function: This is an extension of the standard MMSE spectral amplitude estimator. It introduces a "beta-order" parameter ($\beta$) into the cost function, which allows for a flexible trade-off between minimizing speech distortion and reducing residual noise. This parameter helps to better balance the gains in noise reduction and speech quality, especially for weaker spectral components.
  • Chi Speech Prior: Instead of the traditional Rayleigh distribution, the authors employ a Chi distribution to model the statistical properties of the speech prior. The Chi distribution includes a shape parameter ($\alpha$) that can be adjusted to control the trade-off between residual noise and musical tones. This provides a more generalized and potentially more accurate representation of speech characteristics.
  • Combined Estimator: The paper mathematically combines the beta-order cost function with the Chi speech prior to derive a new closed-form estimator for the spectral amplitude. This new estimator incorporates both the $\beta$ parameter from the cost function and the $\alpha$ parameter from the Chi prior.

The authors also utilize standard signal processing techniques like the Short-Time Fourier Transform (STFT) to analyze the signal in the frequency domain and the overlap-add (OLA) technique for reconstructing the enhanced signal.

Datasets

The evaluation of the proposed estimator was performed using the NOIZEUS corpus. This corpus contains:

  • 30 IEEE sentences spoken by six different speakers (three male, three female).
  • These clean sentences were corrupted by eight different types of real-world noises (e.g., train, babble, car, street, airport).
  • The noise was added at various Signal-to-Noise Ratios (SNRs) ranging from 0 dB to 15 dB in 5 dB increments.

The noises were sourced from the AURORA database.

Key Insights and Ideas

  • Synergy of Flexibility: The paper demonstrates that combining a flexible cost function (beta-order) with a flexible prior model (Chi) leads to superior performance compared to using either improvement in isolation.
  • Chi Prior's Advantage: The Chi prior offers a more nuanced way to model speech characteristics than the simpler Rayleigh distribution, allowing for better adaptation to different noise conditions and speech types.
  • Parameter Tuning is Key: The performance of the estimator is sensitive to the chosen values of the shape parameter ($\alpha$) and the beta-order parameter ($\beta$). Finding the optimal combination of these parameters is crucial for maximizing speech enhancement.
  • Improved Trade-offs: The new estimator effectively manages the inherent trade-offs in speech enhancement, leading to reduced musical noise and speech distortion while simultaneously improving noise reduction.

Key Results

The proposed MMSE Beta-Order Spectral Amplitude estimator with Chi speech prior was evaluated using Segmental Signal-to-Noise Ratio (SSNR) and Perceptual Evaluation of Speech Quality (PESQ) metrics.

  • SSNR Improvements: The new estimator achieved gains of 0-3 dB in SSNR compared to the baseline MMSE Beta-Order Spectral Amplitude estimator using Rayleigh priors. These improvements were most pronounced at lower SNRs (0 dB and 5 dB) and for specific noise types like car and train noises.
  • PESQ Improvements: The estimator provided gains of 0-0.03 in PESQ. While seemingly small, these improvements translate to a noticeable increase in perceived speech quality. At input SNRs of 0 dB and 5 dB, PESQ improvements reached up to 0.55 for car noise.
  • Parameter Impact: The shape parameter $\alpha$ of the Chi prior was shown to influence the trade-off between residual noise and musical tones. The beta-order parameter $\beta$ was fixed at -1 for most experiments, a value found to balance noise reduction and speech quality effectively.

Limitations and Future Work

  • Parameter Optimization: The paper acknowledges that the optimal combination of the shape parameter ($\alpha$) and the beta-order parameter ($\beta$) for the Chi distribution and beta-order cost function was not fully explored.
  • Future Work: The authors suggest developing and implementing an optimal combination of these parameters to achieve even further improvements in noise reduction and overall speech quality. They also propose exploring other statistical models and cost functions.

Summary updated

BibTeX
@inproceedings{trawicki2012improvements,
  author = {Marek Trawicki and Michael Johnson},
  title = {{Improvements of the beta-order minimum mean-square error (MMSE) spectral amplitude estimator using chi priors}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure