Speech enhancement with bivariate gamma model

Atanu Saha, Tetsuya ShimamuraInterspeech 2012

TL;DR

This paper introduces a new family of Bayesian speech enhancement estimators that leverage a bivariate Gamma distribution to model correlated spectral components in speech. Experimental results demonstrate that these novel estimators outperform existing state-of-the-art methods in terms of overall speech quality, particularly in noisy conditions.

What Problem Does This Paper Address?

Speech enhancement aims to improve the quality of noisy speech signals. A common approach involves estimating the clean speech spectrum from the noisy observation. While many successful methods, like the Minimum Mean Square Error (MMSE) estimator, assume that spectral components are uncorrelated, this assumption is often violated in practice. This violation stems from factors like the finite analysis window in signal processing and the inherent harmonic structure of voiced speech. Ignoring these correlations can lead to suboptimal performance and introduce annoying artifacts like musical noise. This paper addresses this limitation by proposing estimators that explicitly model these correlations.

Methods

The core contribution of this paper is the development of a family of Bayesian estimators based on a bivariate Gamma distribution. This distribution is chosen because it can model correlated random variables, which is crucial for capturing the relationships between spectral components of speech.

The authors derive three specific estimators by considering different transformations of two correlated variables modeled by the bivariate Gamma distribution:

  1. Sum of two correlated variables: This leads to the $G_k^S$ gain function.
  2. Product of two correlated variables: This results in the $G_k^P$ gain function.
  3. Ratio of two correlated variables: This yields the $G_k^R$ gain function.

These gain functions are then used to estimate the clean speech spectral amplitude ($\hat{A}_k$) from the noisy observation ($Y_k$) using the relationship $\hat{A}_k = G_k R_k$, where $R_k$ is the noisy spectral amplitude. The estimators are designed to minimize a cost function, and a small power exponent ($\tau$) is used in the cost function to retain the desirable characteristics of the Log-Spectral Amplitude (LSA) estimator, helping to reduce musical noise.

Datasets

The paper uses the NOIZEUS speech corpus for evaluation. This corpus contains speech corrupted by different types of nonstationary noises (exhibition noise and street noise) at various levels (0 dB, 5 dB, 10 dB, 15 dB SNR). The authors state that all utterances from the corpus were used in the experiments.

Key Insights and Ideas

  • Correlated spectral components matter: The paper strongly suggests that modeling the correlation between spectral components is essential for achieving superior speech enhancement performance, moving beyond the limitations of uncorrelated models.
  • Bivariate Gamma distribution as a powerful tool: The bivariate Gamma distribution proves to be a flexible and effective model for capturing these spectral correlations, leading to novel estimator designs.
  • Family of estimators offers flexibility: By deriving estimators based on sum, product, and ratio of correlated variables, the authors provide a versatile set of tools that can be applied in different scenarios.
  • Balancing noise reduction and speech distortion: The shaping parameter ($\nu$) of the Gamma distribution plays a critical role. The authors found that a value of $\nu=0.6$ offers a good compromise between effectively reducing noise and minimizing speech distortion, avoiding the extreme attenuation or distortion seen with very small or large $\nu$ values.
  • Improved perceptual quality: Beyond objective measures, listeners reported that speech processed by the proposed estimators, particularly BGMS, sounds more perceptual due to a more balanced suppression of noise without excessive distortion.

Key Results

The experimental results, summarized in Table 1, show that the proposed estimators (BGMS, BGMP, BGMR) consistently outperform the $\beta$-MMSE estimator across different noise types and levels. The boldface numbers in Table 1 indicate the best performance, and the proposed methods are frequently highlighted as the best performers.

For example, in exhibition noise at 15 dB SNR, BGMS achieves a Coul score of 3.31, compared to 3.17 for $\beta$-MMSE. Similarly, for street noise at 0 dB SNR, BGMS achieves 2.10, outperforming $\beta$-MMSE's 1.99. The paper also notes that the suppression provided by the proposed estimators remains relatively constant at low SNRs, which is beneficial for avoiding exaggerated suppression and subsequent distortion.

Limitations and Future Work

  • Parameter selection: While the paper discusses the importance of the shaping parameter $\nu$ and suggests $\nu=0.6$ for experiments, the optimal selection of parameters for the bivariate Gamma distribution might require further investigation for diverse acoustic conditions.
  • Computational complexity: Not explicitly discussed, but complex statistical models can sometimes lead to higher computational costs, which might be a consideration for real-time applications.
  • Future work: The paper suggests that the proposed family of estimators provides a solid foundation for further research. Potential future directions could include exploring different cost functions, investigating more advanced correlation models, or extending these methods to multi-channel speech enhancement.

Summary updated

BibTeX
@inproceedings{saha2012speech,
  author = {Atanu Saha and Tetsuya Shimamura},
  title = {{Speech enhancement with bivariate gamma model}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure