Abstract
In this paper, a residual echo cancellation method is proposed that uses an estimation of the minimum mean-square error (MMSE) based on a statistical model of a speech signal and an echo signal. After the suppression of the echo signal based on the adaptive filter, residual echo is further reduced by the proposed MMSE estimator and the results are compared with the conventional Wiener filter based method.
Introduction
Hands-free communication systems, such as speakerphones and PC-based chat applications, frequently encounter the challenge of acoustic echo. An Acoustic Echo Canceller (AEC) is typically employed to prevent the far-end signal from looping back through the microphone. While standard AECs utilize adaptive filters to estimate and subtract the echo, these filters often leave behind a "residual echo" due to hardware limitations or dynamic changes in the acoustic environment.
Traditionally, post-filters like the Wiener filter have been used to address this residual echo. However, Wiener filters often introduce "musical noise"—distracting, chirping artifacts caused by rapid fluctuations in the filter's gain. This paper proposes a refined approach using a Minimum Mean-Square Error (MMSE) estimator to achieve superior echo suppression with fewer digital artifacts.
Proposed Methodology
The authors propose a two-stage process for echo cancellation. First, a standard Normalized Least Mean-Square (NLMS) adaptive filter handles the primary echo cancellation. Second, an MMSE short-time spectral amplitude estimator is applied to the remaining signal to treat the residual echo as a statistical problem.
The core of the method lies in statistical modeling, assuming that both the near-end speech and the echo follow a complex Gaussian distribution. This allows for the derivation of a gain function involving modified Bessel functions. To mitigate musical noise, the researchers implemented a "Decision-Directed" (DD) approach to estimate the a priori Signal-to-Echo Ratio (SER). This acts as a temporal smoother, preventing abrupt changes in the filter's gain and ensuring a more natural sound.
Experimental Setup
The algorithm's performance was evaluated using 20 test sentences (10 male, 10 female) from the P.862 standard. A male speaker's sentence served as the far-end echo source. To simulate real-world conditions, "babble" and vehicular noise from the NOISEX-92 database were added at various Signal-to-Noise Ratios (SNR). The audio processing was conducted at an 8 kHz sampling rate with 10 ms frames.
Results and Discussion
The primary advantage of the MMSE estimator over the conventional Wiener filter is the reduction of musical noise. By utilizing the DD approach, the gain applied to the audio signal becomes significantly smoother over time. While raw Echo Return Loss Enhancement (ERLE) measurements showed competitive performance, the qualitative improvement in "naturalness" was the most significant finding.
Treating echo as a statistical noise source—adapting techniques like the Ephraim-Malah algorithm—proves effective for speech enhancement in AEC contexts. The smoother gain function leads to a better user experience by preserving the quality of the background audio while effectively suppressing the echo.
Limitations and Future Work
The current mathematical model simplifies the problem by assuming the absence of background noise during the derivation, although noise was included in the testing phase. Future iterations could incorporate a more robust three-way statistical model (speech + echo + noise). Additionally, the computational complexity of Bessel functions in the MMSE estimator remains a consideration for low-power hardware. Further research could also explore the optimization of the smoothing factor ($\alpha$) across diverse acoustic environments.
Summary updated