Fast converging iterative kalman filtering for speech enhancement using long and overlapped tapered windows with large side lobe attenuation

Stephen So, Kuldip K. PaliwalInterspeech 2010

TL;DR

This paper introduces a novel iterative Kalman filtering approach for speech enhancement that significantly improves convergence speed and reduces residual noise compared to existing methods. By employing long, overlapped frames and a Dolph-Chebychev window for initial linear prediction analysis, the method achieves better speech quality and intelligibility, especially in noisy conditions.

What Problem Does This Paper Address?

Speech enhancement aims to improve the quality and intelligibility of speech signals corrupted by noise. While Kalman filtering offers advantages like inherent speech production modeling and avoidance of musical noise, its performance degrades significantly in low Signal-to-Noise Ratios (SNRs) when parameters are estimated from noisy speech. Existing iterative Kalman filtering methods, while improving performance over a few iterations, can introduce undesirable "musical noise" and speech distortion. This paper addresses the challenge of achieving faster convergence and cleaner enhanced speech from iterative Kalman filters, particularly in low SNR scenarios.

Methods

The core of this paper is an enhanced iterative Kalman filtering scheme. The key innovations are:

  • Long and Overlapped Frames: Instead of short, non-overlapping frames (typically 20ms), the authors use longer (80ms) and overlapped frames (10ms update). This provides more statistically reliable autocorrelation estimates for linear prediction and ensures frequent updates of model parameters, leading to better temporal smoothing.
  • Tapered Window with Large Side Lobe Attenuation: For the initial linear prediction (LPC) analysis in the first iteration, a Dolph-Chebychev window with a -200 dB side lobe attenuation is used. This window is chosen for its ability to enhance the dynamic range of speech formants, reduce bias in prediction error variance, and provide spectral smoothing. In subsequent iterations, a rectangular window is used.
  • Iterative Refinement: The process involves an initial Kalman filtering step using the enhanced window and frame settings, followed by re-estimation of LPCs and excitation variances from the enhanced speech. This iterative process is repeated for a few cycles.

Datasets

The experiments were conducted using the NOIZEUS speech corpus. This corpus consists of 30 phonetically balanced sentences from six speakers, sampled at 8 kHz. For evaluation, a stimuli set was created by adding additive white Gaussian noise to the clean speech at four different SNR levels: 0, 5, 10, and 15 dB.

Key Insights and Ideas

  • Window Choice Matters for Initial Estimates: The choice of window for the initial LPC analysis is crucial. A window with high side lobe attenuation (like Dolph-Chebychev) can significantly improve the initial estimates, leading to better overall performance and faster convergence of the iterative Kalman filter.
  • Longer Frames Improve Robustness: Using longer, overlapped frames provides more stable estimates of speech parameters, making the Kalman filter more robust to noise.
  • Reduced Musical Noise and Distortion: The proposed method effectively reduces the annoying "musical noise" often associated with iterative Kalman filters. While it might introduce a slight "washy" residual artifact, this is considered less bothersome than musical noise.
  • Faster Convergence: The proposed method achieves better performance with fewer iterations compared to previous iterative Kalman filters, indicating faster convergence.

Key Results

  • Improved PESQ Scores: Objective evaluations using the Perceptual Evaluation of Speech Quality (PESQ) measure showed that the proposed iterative Kalman filter consistently outperformed conventional iterative and non-iterative Kalman filters, as well as MMSE-STSA and PSC methods across various SNRs.
  • Convergence Achieved in Two Iterations: The proposed method achieved near-optimal performance with just two iterations, whereas the iterative Kalman filter from Gibson et al. [5] required up to three iterations for comparable or slightly lower scores.
  • Qualitative Improvement: Spectrogram analysis revealed significantly less wideband residual noise in the proposed method's output compared to other methods. Informal listening tests confirmed the absence of musical tones and a less annoying residual artifact.

Limitations and Future Work

  • Oversuppression of Speech: The paper notes that the proposed method can lead to oversuppression of speech, which might affect intelligibility. This is an area for potential improvement.
  • "Washy" Residual Artifact: While less annoying than musical noise, the "washy" residual artifact observed in informal listening tests suggests that further refinement might be possible to achieve even cleaner speech.

Future work could focus on mitigating the speech oversuppression and further refining the residual artifact to achieve even higher quality speech enhancement.

Summary updated

BibTeX
@inproceedings{so2010fast,
  author = {Stephen So and Kuldip K. Paliwal},
  title = {{Fast converging iterative kalman filtering for speech enhancement using long and overlapped tapered windows with large side lobe attenuation}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure