Non-stationary noise power spectral density estimation based on regional statistics

Xiaofei Li, Laurent Girin, Sharon Gannot, Radu Horaud

TL;DR

This paper introduces a novel method for estimating non-stationary noise power spectral density (PSD) by leveraging "regional statistics" derived from short-term periodogram analysis. This approach improves the accuracy of speech enhancement algorithms by better tracking rapidly changing noise levels, even in challenging scenarios.

What Problem Does This Paper Address?

Accurate estimation of the noise power spectral density (PSD) is a cornerstone of effective single-channel speech enhancement. Traditional methods often struggle with non-stationary noise – noise whose characteristics change over time. Many existing algorithms rely on finding the minimum spectral power in a time-frequency window, which can lead to overestimation if speech is present or significant delays in tracking noise increases. Other methods, while avoiding minimum tracking, can still be slow to adapt to abrupt noise level changes. This paper aims to bridge this gap by developing a noise PSD estimation technique that is both robust to non-stationary noise and capable of rapid adaptation.

Methods

The core innovation of this paper lies in the introduction of "regional statistics." These are four features computed from the periodograms of the noisy speech signal over a short time window (approximately 0.2 seconds). These features are designed to capture the statistical differences between noise-only periods and periods containing speech. The four regional statistics are:

  • Normalized Variance ($\theta_{nv}$): Measures the variance of past periodograms relative to their smoothed average. Higher values suggest speech presence due to abrupt changes.
  • Normalized Differential Variance ($\theta_{ndv}$): Captures the variance of the difference between consecutive periodograms. This helps distinguish between correlated speech segments and uncorrelated noise.
  • Normalized Average Variance ($\theta_{nav}$): Assesses the variance of the average of consecutive periodograms. This is shown to be a more distinctive feature than the normalized variance alone.
  • Median Crossing Rate ($V_{mcr}$): Tracks how often the periodogram crosses its median value. This rate is expected to be lower for speech due to frame correlation.

These regional statistics are then used to estimate the Speech Presence Probability (SPP). The SPP is a crucial component that controls a time-varying smoothing parameter in the recursive noise PSD estimation. A higher SPP indicates a higher likelihood of speech, leading to a more cautious update of the noise PSD estimate, while a lower SPP allows for faster adaptation to changes. The paper also introduces a normalized distance metric based on these statistics to further aid in the SPP estimation. The noise PSD is then recursively updated using a weighted average, where the weights are dynamically adjusted by the SPP.

Datasets

The paper uses the TIMIT database for clean speech signals. For noise, it employs:

  • Computer-generated stationary White Gaussian Noise (WGN).
  • Non-stationary WGN (generated by fluctuating the variance of stationary WGN).
  • Factory noise.
  • Babble noise (from the NOISEX92 database).

The experiments evaluate performance across various Signal-to-Noise Ratios (SNRs) from 0 dB to 15 dB.

Key Insights and Ideas

  • Regional Statistics as Discriminators: The paper convincingly demonstrates that by analyzing the statistical behavior of periodograms over a short, localized time window, one can effectively differentiate between noise-only and noisy speech segments. This is a more nuanced approach than simply looking for spectral minima.
  • SPP as a Control Mechanism: Using the estimated SPP to dynamically adjust the noise PSD smoothing parameter is a key insight. This allows the algorithm to be conservative when speech is likely present (preserving speech details) and aggressive when noise is changing rapidly (tracking the noise).
  • Robustness to Abrupt Changes: The proposed method shows a significant advantage in tracking noise PSDs that increase abruptly (e.g., 10 dB/s), outperforming existing methods like IMCRA and MMSE-based estimators that exhibit longer adaptation delays.
  • Trade-off Between Noise and Speech: The regional statistics are designed to exploit the inherent differences in the temporal correlation and variability between speech and noise signals.

Key Results

  • The proposed method achieves good tracking performance for non-stationary noise, even with abrupt increases in noise level.
  • In terms of the symmetric segmental logarithmic error (LogErr), the proposed method generally performs better or comparably to IMCRA and MMSE-based methods across various noise types and SNRs.
  • For non-stationary noise, the proposed method shows a notable improvement in LogErr and output SNR compared to IMCRA.
  • The paper highlights that the regional statistics contribute to the discrimination between noise and speech, with all four features showing similar usefulness when removed individually.

Limitations and Future Work

  • Assumption of Noise Stationarity: The validity of the regional statistics relies on the assumption that the noise PSD is stationary within the short time window used for computation. While the method performs well even with abrupt noise changes, extreme non-stationarity within this window could still pose challenges.
  • Inference of Statistics: The paper infers the expectation vector and covariance matrix for regional statistics from stationary WGN. While this works well, the performance might be further improved if these statistics could be adapted or learned for different types of noise.
  • Future Work: The authors suggest that the proposed method's ability to efficiently track non-stationary noise makes it suitable for various speech enhancement applications. Further exploration into adapting the regional statistics for more diverse noise environments could be a promising direction.

Summary updated

Abstract

Estimating the noise power spectral density (PSD) is essential for single channel speech enhancement algorithms. In this paper, we propose a noise PSD estimation approach based on regional statistics. The proposed regional statistics consist of four features representing the statistics of the past and present periodograms in a short-time period. We show that these features are efficient in characterizing the statistical difference between noise PSD and noisy speech PSD. We therefore propose to use these features for estimating the speech presence probability (SPP). The noise PSD is recursively estimated by averaging past spectral power values with a time-varying smoothing parameter controlled by the SPP. The proposed method exhibits good tracking capability for non-stationary noise, even for abruptly increasing noise level.
BibTeX
@inproceedings{li2016non,
  author = {Xiaofei Li and Laurent Girin and Sharon Gannot and Radu Horaud},
  title = {{Non-stationary noise power spectral density estimation based on regional statistics}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2016},
  url = {https://openreview.net/forum?id=L8gf36EF3C},
}
Figure