Adaptive filtering for attenuating musical noise caused by spectral subtraction

Takahiro Murakami, Yoshihisa IshidaInterspeech 2006

TL;DR

This paper tackles the "musical noise" artifact introduced by spectral subtraction, a common speech enhancement technique. They propose using an adaptive filter, leveraging the differing durations of speech and musical noise components, to effectively remove this annoying distortion and improve overall speech quality.

What Problem Does This Paper Address?

Speech enhancement techniques are crucial for improving the usability of speech processing applications in real-world, noisy environments. Spectral subtraction is a popular method for noise reduction due to its simplicity and low computational cost. However, when applied to noisy speech, especially at low signal-to-noise ratios (SNRs), it can introduce an undesirable artifact known as "musical noise." This noise manifests as isolated, short-duration frequency components that sound like random musical tones, significantly degrading the intelligibility and perceived quality of the enhanced speech. The paper aims to address this specific distortion problem and improve the quality of speech enhanced by spectral subtraction.

Methods

The core of the proposed method is the application of an adaptive filter after the initial spectral subtraction. Here's how it works:

  1. Spectral Subtraction: The noisy speech is first processed using a conventional spectral subtraction algorithm. This step aims to remove the background noise but, as noted, introduces musical noise.
  2. Adaptive Filtering for Musical Noise Reduction: The output of the spectral subtraction (the "enhanced" speech, which still contains musical noise) is then fed into an adaptive filter. Crucially, this enhanced speech is used as the reference signal for the adaptive filter. The original noisy speech is used as the input signal to the adaptive filter.
  3. Exploiting Temporal Differences: The key insight is the difference in duration between speech components and musical noise components. Musical noise is characterized by short-lived frequency bursts (typically < 20ms), while speech components are generally longer-lasting. The adaptive filter is configured with a convergence speed that is faster than the typical duration of speech components but slower than the typical duration of musical noise components.
  4. Mechanism: Because the adaptive filter's convergence speed is tuned this way, it can effectively "track" and follow the longer-duration speech components. However, it cannot adapt quickly enough to the transient, short-lived musical noise components. As a result, the musical noise is attenuated in the filter's output, while the speech signal is preserved. The paper specifically mentions using the Normalized Least-Mean-Square (N-LMS) algorithm for the adaptive filter.

Datasets

Not discussed in the paper. The paper mentions using speech uttered by "two male and two female speakers" for simulation, but no specific dataset names are provided.

Key Insights and Ideas

  • Musical Noise as a Temporal Artifact: The paper frames musical noise not just as a spectral artifact but as one with distinct temporal characteristics (short duration) that can be exploited.
  • Adaptive Filter as a "Temporal Gate": By carefully tuning the adaptive filter's convergence speed, it acts as a temporal gate, allowing longer-duration speech to pass through while blocking shorter-duration musical noise.
  • Leveraging Spectral Subtraction Output: Instead of trying to fix spectral subtraction's flaws directly in the spectral domain, the paper cleverly uses its output as a reference for a secondary, adaptive filtering stage.
  • Trade-off Between SNR and Cepstral Distance: The simulations reveal an interesting trade-off: aggressive musical noise reduction (lower cepstral distance) can sometimes slightly degrade the overall SNR, and vice-versa. This highlights the need for careful parameter tuning.

Key Results

  • The proposed method effectively reduces musical noise, leading to improved speech enhancement quality compared to standard spectral subtraction (SS) and spectral subtraction with conventional post-processing (SS+PP).
  • Cepstral Distance: The proposed method, particularly with a step-size parameter $\mu = 0.1$ (Proposed1), significantly reduces the averaged cepstral distance, which is an indicator of intelligibility.
  • SNR: While the lowest cepstral distance case (Proposed1) showed a slight degradation in SNR, other configurations (e.g., Proposed2 with $\mu = 0.5$) maintained a good SNR while still achieving substantial cepstral distance reduction.
  • The paper suggests that by adjusting the step-size parameter ($\mu$) of the N-LMS algorithm, the enhancement quality can be efficiently improved, balancing musical noise reduction and SNR preservation.

Limitations and Future Work

  • Parameter Tuning: The paper acknowledges that determining the "theoretically reasonable step-size parameter" for the adaptive filter is a direction for future work. The current parameter settings were found empirically.
  • Manual Speech Activity Detection: The determination of speech activity segments for the spectral subtraction part was done by manual inspection, which is not practical for real-time systems.
  • Generalization: While the paper shows good results, further investigation into the performance across a wider variety of noise types and speech conditions might be beneficial.

Summary updated

BibTeX
@inproceedings{murakami2006adaptive,
  author = {Takahiro Murakami and Yoshihisa Ishida},
  title = {{Adaptive filtering for attenuating musical noise caused by spectral subtraction}},
  booktitle = {Interspeech},
  year = {2006},
}
Figure