Estimation of relative transfer function in the presence of stationary noise based on segmental power spectral density matrix subtraction

Xiaofei Li, Laurent Girin, Radu Horaud, Sharon Gannot

TL;DR

This paper introduces a novel method for estimating the relative transfer function (RTF) of a speech source in the presence of stationary noise. By subtracting power spectral density (PSD) matrices from speech-dominated and noise-dominated segments, the authors effectively remove noise, leading to more accurate RTF estimation, particularly beneficial for sound source localization.

What Problem Does This Paper Address?

Estimating the relative transfer function (RTF) is crucial for various audio processing tasks like beamforming and sound source localization. The challenge arises when stationary noise contaminates the desired speech signal. Existing methods often struggle with this noise, either by relying on accurate noise PSD estimation or by processing each channel pair independently, which can be suboptimal. This paper aims to provide a robust RTF estimation technique that effectively handles stationary noise by leveraging the non-stationary nature of speech.

Methods

The core of the proposed method is segmental power spectral density (PSD) matrix subtraction. Here's a breakdown:

  1. Segmentation: The audio signal from a microphone array is divided into segments. These segments are then classified into two types: those with high speech power (speech-plus-noise) and those with low speech power (noise-only).
  2. PSD Matrix Calculation: For each segment, a PSD matrix is computed. This matrix captures the spectral characteristics of the signals within that segment.
  3. Noise Subtraction: The key innovation is subtracting the PSD matrix of a noise-dominated segment from the PSD matrix of a speech-dominated segment. Since the noise is assumed to be stationary, its contribution to the PSD matrix is relatively consistent. By subtracting, the stationary noise component is significantly reduced, while the non-stationary speech component is preserved.
  4. RTF Estimation: The resulting "noise-free" PSD matrix is then used for RTF identification. This is achieved by computing the principal eigenvector of this matrix, which provides a robust estimate of the RTF for a single speaker.
  5. Segment Classification: To classify segments, the authors utilize minimum and maximum statistics of the noise signal. They define thresholds based on the ratio of maximum to minimum noise power within segments to distinguish between speech-heavy and noise-heavy segments. This approach is designed to be robust to variations in segment increment size.

Datasets

  • Training Data: An acoustic-visual dataset was used for training. This dataset consisted of 1-second white-noise signals emitted from 432 source directions within a 24° × 18° field of view. Ground truth directions were obtained by localizing the sound source's visual marker.
  • Test Data: 108 speech signals from the TIMIT dataset were used. These signals were emitted from 108 different directions within the camera's field of view.
  • Noise: White Gaussian Noise (WGN) and babble noise from the NOISE92 library were used as stationary noise sources. These were added to the speech signals at various signal-to-noise ratios (SNRs).

The paper does not explicitly mention introducing new benchmarks, but rather uses existing datasets for evaluation.

Key Insights and Ideas

  • Leveraging Non-stationarity: The fundamental idea is to exploit the inherent difference in stationarity between speech (non-stationary) and noise (stationary). This difference is the key to effectively separating them.
  • PSD Matrix Subtraction as Noise Reduction: Instead of relying on complex noise PSD estimation, directly subtracting PSD matrices from different segment types offers a more straightforward and effective way to cancel stationary noise.
  • Eigenvector for Robust RTF: Using the principal eigenvector of the "noise-free" PSD matrix provides a globally optimal RTF estimate across all channels and frequencies, leading to improved robustness.
  • Minimum/Maximum Statistics for Classification: The use of minimum and maximum statistics for segment classification is a clever way to adaptively identify segments with significant speech content, making the method less sensitive to segment overlap and increment size.

Key Results

The proposed RTF estimation method was evaluated in the context of sound source localization (SSL).

  • Superior Localization Performance: In experiments with WGN noise, the proposed method consistently achieved better localization accuracy (smaller error in pixels) compared to two existing methods: the non-stationarity (NS) method and the speech presence probability (SPP) method, across all tested SNRs.
  • Robustness to Noise Types: While the proposed method performed best with WGN, it also showed competitive results with babble noise, although babble noise itself introduced larger localization errors due to its relative non-stationarity.
  • Comparison with SPP: The paper highlights that the proposed method's advantage over the SPP method stems from its accurate noise PSD subtraction, whereas SPP relies on the accuracy of noise PSD estimation. Additionally, the eigenvalue decomposition used in the proposed method considers all channels simultaneously, unlike the channel-pair-wise optimization of SPP.

Limitations and Future Work

  • Single Source Assumption: The current method is designed for a single desired speech source. The authors acknowledge that extending it to multiple desired sources is a significant challenge. In such cases, the "desired source" term in the PSD matrix would become a sum of multiple sources, making individual RTF estimation non-straightforward.
  • Future Work: The authors suggest that the proposed PSD matrix subtraction method could still be valuable for multi-source applications, particularly in sound source localization based on subspace methods. Future work will focus on addressing the multi-source scenario.

Summary updated

Abstract

This paper addresses the problem of relative transfer function (RTF) estimation in the presence of stationary noise. We propose an RTF identification method based on segmental power spectral density (PSD) matrix subtraction. First multiple channel microphone signals are divided into segments corresponding to speech-plus-noise activity and noise-only. Then, the subtraction of two segmental PSD matrices leads to an almost noise-free PSD matrix by reducing the stationary noise component and preserving non-stationary speech component. This noise-free PSD matrix is used for single speaker RTF identification by eigenvalue decomposition. Experiments are performed in the context of sound source localization to evaluate the efficiency of the proposed method.
BibTeX
@inproceedings{li2015estimation,
  author = {Xiaofei Li and Laurent Girin and Radu Horaud and Sharon Gannot},
  title = {{Estimation of relative transfer function in the presence of stationary noise based on segmental power spectral density matrix subtraction}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2015},
  url = {https://openreview.net/forum?id=naIMSSqIP4t},
}
Figure