Developments of a hybrid pre-processor based on frequency shifting for stereophonic acoustic echo cancellation

Bruno C. Bispo, Diamantino S. FreitasInterspeech 2012

TL;DR

This paper introduces a novel hybrid approach for stereophonic acoustic echo cancellation (SAEC) that combines frequency shifting with a half-wave rectifier technique. By applying frequency shifts in specific sub-bands of the audio spectrum, the authors significantly improve the accuracy of echo path estimation and the perceptual quality of the processed signals compared to existing methods, while maintaining low computational complexity.

What Problem Does This Paper Address?

In hands-free communication systems, acoustic coupling between loudspeakers and microphones creates an echo that needs to be eliminated. Stereophonic acoustic echo cancellation (SAEC) aims to remove this echo from two audio channels. A key challenge in SAEC is the "non-uniqueness problem," where adaptive filters can converge to incorrect solutions due to strong correlations between the left and right audio channels, especially when they originate from the same sound source. This bias in filter coefficients degrades the echo cancellation performance. To combat this, pre-processing techniques are employed to decorrelate the channel signals before they are fed into the adaptive filters. However, these techniques must be computationally efficient and, crucially, must not negatively impact the perceived quality or spatial location of the audio. Existing methods like adding low-level noise or using half-wave rectification have limitations, particularly in decorrelating higher frequency components without affecting stereo perception.

Methods

The paper proposes a hybrid pre-processing technique for SAEC that combines the strengths of two existing methods: the half-wave rectifier (hwr) and frequency shifting (fs).

  • Half-Wave Rectifier (hwr): This technique introduces non-linearities by adding a positive half-wave rectified version of the signal to one channel and a negative version to the other. While effective for decorrelation, it can struggle with higher frequencies.
  • Frequency Shifting (fs): This method shifts the frequency of one channel relative to the other. While it offers strong decorrelation, applying it to the entire spectrum was found to severely damage stereo perception.
  • Proposed Hybrid Approach: The core innovation is to apply these techniques selectively across different frequency sub-bands. The authors use a filter bank to divide the audio spectrum (e.g., 0-2 kHz, 2-4 kHz, 4-8 kHz). The hwr method is applied in lower frequency bands where it performs well and is perceptually benign. In higher frequency bands, where hwr is less effective, frequency shifting is applied with carefully chosen shift values (wo). This allows for improved decorrelation in the higher frequencies without the drastic perceptual degradation seen with full-spectrum shifting. The paper specifically evaluates two hybrid configurations: hybrid1 and hybrid2, differing in the frequency bands where fs is applied.

Datasets

The paper uses a speech database consisting of 100 signals from 10 different talkers, each with a duration of 4 seconds, originally sampled at 48 kHz and downsampled to 16 kHz for the simulations. This database was used for estimating coherence functions. For experiments involving adaptive filters and subjective listening tests, longer concatenated signals (40 seconds) were used, and a subset of 5 English signals was selected for the subjective evaluation.

Key Insights and Ideas

  • Sub-band processing is key: The success of the hybrid approach hinges on applying different decorrelation strategies to different frequency bands, leveraging the strengths of each method where they are most effective and perceptually acceptable.
  • Frequency shifting's perceptual impact is frequency-dependent: While full-spectrum frequency shifting is disastrous for stereo perception, applying small shifts in higher frequency bands is perceptually tolerable and highly effective for decorrelation. This is linked to the human auditory system's reduced sensitivity to interaural time differences at higher frequencies.
  • Hybrid methods offer a superior trade-off: The proposed hybrid techniques (hybrid1 and hybrid2) demonstrate a significant improvement in echo path estimation and signal quality over the standalone hwr method, without the perceptual drawbacks of full-spectrum fs.

Key Results

  • Improved Decorrelation: Figures 5(b)-(d) show that the hybrid methods (especially hybrid2) achieve significantly lower coherence values in higher frequency bands compared to the hwr method (Figure 5(a)), indicating better decorrelation.
  • Enhanced Echo Path Estimation: Figure 6 demonstrates that the hybrid methods lead to lower normalized misalignment for the adaptive filters. The hybrid2 method achieved a 4 dB gain compared to the hwr method, meaning its estimated echo paths were closer to the true impulse responses.
  • Better Perceptual Quality: Subjective listening tests (Figure 7) showed that the hybrid methods received higher quality grades than the hwr method. The authors attribute this to the fact that frequency shifts in higher frequencies cause less perceptible degradation than the distortions introduced by hwr in those same bands.

Limitations and Future Work

  • Limited Scope of Frequency Shifting: The paper focuses on a hybrid approach that still relies on hwr in lower bands. The authors suggest that further investigation is necessary to develop a decorrelation method using only frequency shifts, potentially across all bands with carefully controlled parameters.
  • Subjective Test Variance: The authors acknowledge that the subjective listening test results showed some variance in the 95% confidence intervals, partly due to the subjective nature of the test and the limited number of evaluators.
  • Computational Complexity: While the proposed methods are described as having low complexity, a detailed analysis of the computational cost of the filter bank and sub-band processing is not explicitly provided.

The paper concludes by highlighting the promise of sub-band frequency shifting for SAEC, offering a path towards improved performance with minimal perceptual impact.

Summary updated

BibTeX
@inproceedings{bispo2012developments,
  author = {Bruno C. Bispo and Diamantino S. Freitas},
  title = {{Developments of a hybrid pre-processor based on frequency shifting for stereophonic acoustic echo cancellation}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure