Mask estimation in non-stationary noise environments for missing feature based robust speech recognition

Shirin Badiezadegan, Richard C. RoseInterspeech 2010

TL;DR

This paper introduces a novel "Speech Presence Probability" (SPP) mask estimation technique for robust speech recognition in noisy, non-stationary environments. By dynamically estimating speech and noise presence without prior knowledge, it significantly improves accuracy compared to traditional methods, especially in challenging babble noise conditions.

What Problem Does This Paper Address?

Automatic Speech Recognition (ASR) systems often struggle in real-world scenarios where background noise is not constant but fluctuates rapidly (non-stationary). Traditional ASR techniques that rely on clean speech training data degrade significantly in such noisy conditions. While "missing feature" approaches aim to mitigate this by identifying and reconstructing corrupted parts of the speech signal using spectro-temporal masks, accurately estimating these masks in non-stationary noise remains a significant challenge. Existing methods often assume stationary noise or require prior knowledge of the noise type and Signal-to-Noise Ratio (SNR), limiting their applicability in dynamic environments.

Methods

The paper proposes a Speech Presence Probability (SPP) mask estimation method. This approach uses a hypothesis test to dynamically detect the presence of speech in time-frequency bins. It continuously updates estimates of background noise levels, SNR, and speech presence probability without needing any prior information about the noise characteristics. This is a departure from traditional Gaussian Mixture Model (GMM) based soft mask estimation, which relies on pre-trained models of speech and noise, assuming knowledge of noise type and SNR.

The SPP mask is then used in a Minimum Mean Squared Error (MMSE) based data imputation framework. This framework reconstructs the missing or unreliable spectral components of the speech signal based on the estimated mask. The core idea is to replace corrupted spectral bins with estimated values that minimize the mean squared error, thereby allowing a clean-trained ASR system to process the "repaired" speech signal.

Datasets

The study evaluates its proposed method on the Aurora 2 speech corpus, specifically using the connected digit task domain. The experiments utilize test set "a" with two types of noise: subway noise (considered more stationary) and speech babble noise (considered non-stationary). The ASR system itself was trained on the clean speech training set of the Aurora 2 corpus.

Key Insights and Ideas

  • Dynamic Masking is Crucial for Non-Stationary Noise: The SPP mask's ability to adapt to rapidly changing noise conditions is key to its success, especially in environments like babble noise.
  • No Prior Knowledge is a Strength: The SPP approach's independence from prior knowledge of noise type or SNR makes it more versatile and applicable to a wider range of real-world scenarios.
  • Mask Accuracy Directly Impacts ASR Performance: The paper reinforces that highly accurate spectro-temporal masks, even if achieved with "ideal" conditions, lead to remarkably high ASR word accuracy, demonstrating the potential of missing feature techniques.
  • Babble Noise Highlights SPP's Advantage: The experimental results clearly show that the SPP mask significantly outperforms the GMM mask in babble noise, suggesting that its adaptive nature is particularly beneficial for non-stationary acoustic environments.

Key Results

The paper presents quantitative results in terms of Automatic Speech Recognition (ASR) Word Accuracy (WAC) across various SNRs (5 dB to 20 dB) for both subway and babble noise conditions.

  • Baseline: The ASR system without any noise compensation performed poorly, especially at lower SNRs (e.g., 19.44% at 5 dB for babble noise).
  • Ideal Mask: Using an "ideal" mask (derived from separate clean and noise files) achieved very high WAC, often exceeding 95% even at 15 dB SNR, demonstrating the effectiveness of missing feature imputation.
  • GMM Mask: The GMM mask, which requires prior knowledge of noise type and SNR, showed moderate improvements over the baseline but was outperformed by the SPP mask in babble noise. For instance, at 10 dB babble noise, GMM mask achieved 69.95% WAC, while SPP achieved 89.51%.
  • SPP Mask: The proposed SPP mask consistently outperformed the GMM mask in babble noise across all tested SNRs. For subway noise, the performance was comparable, with SPP showing a slight edge at the lowest SNR (5 dB).

Limitations and Future Work

  • Limitations: The paper acknowledges that the Aurora 2 corpus, while useful, has somewhat limited conditions regarding constant average SNR for each utterance. The GMM mask implementation also relies on prior knowledge of noise type and SNR, which is a limitation in truly unknown environments.
  • Future Work: The authors suggest that the SPP approach's ability to continually update estimates of background level, SNR, and speech presence probability is a significant advantage. Further exploration of this adaptive masking technique in more diverse and challenging acoustic environments would be a logical next step.

Summary updated

BibTeX
@inproceedings{badiezadegan2010mask,
  author = {Shirin Badiezadegan and Richard C. Rose},
  title = {{Mask estimation in non-stationary noise environments for missing feature based robust speech recognition}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure