Abstract
The relative transfer function generalized sidelobe canceler (RTF-GSC) is a popular method for implementing multichannel speech enhancement. However, an accurate estimation of channel transfer function ratios poses a challenge, especially in noisy environments. In this work, we demonstrate that even a very low-order RTF estimate can give superior performance in terms of noise reduction without incurring excessive speech distortion. We show that noise reduction is dependent on the correlation between the input noise and the noise reference generated by the Blocking Matrix (BM), and that a low-order RTF estimate preserves this correlation better than a high-order one. The performance for both high-order and low-order RTF estimates are compared using output SNR, Noise Reduction, and a perceptual measure for speech quality.
Introduction
In multichannel speech enhancement, the Generalized Sidelobe Canceler (GSC) is a standard framework. It works by splitting the task into three parts: a fixed beamformer (FBF) to grab the speech, a blocking matrix (BM) to create a "noise-only" reference, and an adaptive noise canceler (ANC) to subtract that noise from the speech signal.
The effectiveness of this system relies heavily on the Relative Transfer Function (RTF), which describes how sound travels between different microphones. Traditionally, researchers assumed that a high-order FIR filter (one with many "taps" or coefficients) was necessary to accurately model these complex acoustic paths. However, estimating these high-order filters in noisy, real-world environments is incredibly difficult and often leads to poor performance. This paper asks: do we actually need that complexity for effective noise reduction?
Methodology
The authors analyze the RTF-GSC algorithm, focusing specifically on how the order of the RTF estimate affects the final stage: the Adaptive Noise Canceler.
- Mathematical Analysis: They derive the relationship between output noise power and the correlation between the FBF and BM outputs. They prove that for diffuse white noise, this correlation is a monotonically decreasing function of the RTF filter order. In simpler terms: the more complex your filter, the less correlated your noise references become, which hurts the cancellation process.
- Comparison: They compared two extremes:
- $RTF_{high}$: A high-order FIR filter with 300 coefficients.
- $RTF_{low}$: A minimal FIR filter with just one non-zero coefficient (representing a simple delay and scaling factor).
- Implementation: The system was tested using a 5-microphone array in a simulated room environment (5m x 4m x 2.8m) with a reverberation time ($T_{60}$) of 0.2 seconds. The ANC stage used a standard Normalized Least Mean Square (NLMS) algorithm, but adaptation was restricted to "silent" intervals to prevent speech distortion.
Key Results
The $RTF_{low}$ approach consistently outperformed the $RTF_{high}$ approach across several metrics:
- Output SNR: $RTF_{low}$ provided a higher Signal-to-Noise Ratio, especially in very noisy conditions (input SNR < 0dB).
- Noise Reduction (NR): The low-order estimate achieved significantly better noise suppression, roughly 4-6 dB better than the high-order version at low input SNRs.
- Speech Quality: Using a Mean Opinion Score (MOS) scale, the researchers found that $RTF_{low}$ maintained "fairly natural" speech quality, performing slightly better than $RTF_{high}$ at lower input SNRs and matching it at higher ones.
Discussion and Conclusion
The "aha moment" of the paper is the realization that the ANC's performance depends entirely on the correlation between the noise in the speech branch and the noise in the reference branch. While a high-order RTF might theoretically model the acoustic path more accurately, it inadvertently "decorrelates" the noise components. A low-order estimate preserves this crucial correlation much better.
Furthermore, a low-order RTF is significantly easier to estimate and requires far less processing power, making it more practical for real-time applications like hearing aids or mobile phones. Surprisingly, using a simpler RTF didn't significantly degrade speech quality. Because the adaptive filter only updated during noise-only periods, the "leakage" of speech into the noise reference was effectively managed.
Future work will address limitations regarding non-stationary noise environments and directional noise sources, as well as the system's dependency on robust Voice Activity Detection (VAD).
Summary updated