Speech Compressive Sampling Using Approximate Message Passing and a Markov Chain Prior

Xiaoli Jia, Peilin Liu, Sumxin JiangSensors 2020 · OpenReview

TL;DR

This paper tackles the challenge of reconstructing speech signals from compressed samples, which are inherently not perfectly sparse. They introduce a novel method combining Approximate Message Passing (AMP) with a Markov chain prior to exploit the temporal and frequency correlations within speech spectrograms, achieving improved reconstruction quality and speech enhancement.

What Problem Does This Paper Address?

Compressive Sampling (CS) allows for signal recovery from fewer samples than traditionally required by the Nyquist-Shannon theorem, but it relies on the signal being sparse in some domain. Speech signals, while somewhat compressible, are not perfectly sparse in standard bases like MDCT. This "near-sparsity" makes traditional CS methods struggle to achieve satisfactory reconstruction, especially when dealing with the complex nature of voiced and unvoiced sounds. The paper aims to bridge this gap by developing a CS method that better leverages the inherent structure of speech signals.

Methods

The core of the proposed method is a turbo message passing framework that iteratively combines two key components:

  • Approximate Message Passing (AMP): This algorithm is used to estimate the marginal distribution of Modified Discrete Cosine Transform (MDCT) coefficients within each speech frame. AMP is chosen for its efficiency in handling sparse signals from compressed measurements.
  • Markov Chain Prior: The paper models the temporal correlation of voiced sounds and the frequency-axis correlation of unvoiced sounds using a Markov chain. Specifically, the "support" (indicating whether a coefficient is non-zero) of MDCT coefficients across frames (for voiced sounds) and across frequency bins (for unvoiced sounds) is modeled as a Markov chain. This captures the inherent structure and dependencies within the speech spectrogram.

The turbo framework alternates between AMP (for observation sector) and belief propagation along the Markov chain (for pattern sector). A crucial addition is a stopping condition for the turbo iterations, designed to prevent divergence and ensure reliable recovery. The method also incorporates a Bernoulli-Gaussian mixture model for MDCT coefficients and uses an Expectation-Maximization (EM) algorithm to learn model parameters.

Datasets

The paper evaluates its method using speech excerpts from:

  • TIMIT acoustic-phonetic continuous speech corpus: Used for adult male and female voices.
  • NOIZEUS database: Used for noisy speech signals (babble noise) for speech enhancement experiments.

The authors also mention reconstructing speech from cellphone recordings.

Key Insights and Ideas

  • Speech Spectrogram Structure is Key: The paper highlights that speech isn't just "nearly sparse" but has exploitable temporal (voiced sounds) and frequency (unvoiced sounds) structures that can be leveraged for better CS reconstruction.
  • Turbo Iteration for Structure Exploitation: Combining AMP with belief propagation on a Markov chain allows for a powerful iterative refinement process that progressively incorporates signal structure.
  • Markov Chain for Support Dynamics: Modeling the support of MDCT coefficients as a Markov chain effectively captures the slow, dynamic changes in sparsity patterns over time and frequency.
  • Stopping Condition Prevents Divergence: The proposed stopping condition is critical for stabilizing the turbo iteration process, preventing over-smoothing and ensuring accurate recovery.
  • Speech Enhancement as a Bonus: The method not only reconstructs clean speech but also demonstrates a comparable speech enhancement effect to state-of-the-art methods.

Key Results

  • Improved SNR: The proposed method consistently achieved higher Signal-to-Noise Ratios (SNR) compared to traditional CS methods like DCS-AMP, EM-GM-AMP, MRF, and FISTA, especially at higher undersampling ratios.
  • Better Spectrogram Similarity: Reconstructed speech spectrograms from the proposed method were more similar to the original spectrograms than those from other algorithms, indicating better preservation of spectral details.
  • Higher PESQ Scores: The method achieved the highest Perceptual Evaluation of Speech Quality (PESQ) scores, indicating superior subjective speech quality. For example, for male's voice, it achieved 3.0974, compared to DCS-AMP's 2.8834 and FISTA's 2.9226.
  • Comparable Speech Enhancement: In speech enhancement tasks, the proposed method achieved comparable PESQ and STOI (Short-Time Objective Intelligibility) scores to a state-of-the-art reference method.

Limitations and Future Work

  • Common Support Assumption: The paper notes that while their method models temporal correlation, the assumption of common support (as in some MMV problems) is not always true for speech. Their dynamic CS approach addresses this by modeling support changes.
  • Amplitude Modeling: The authors mention that they did not exploit the structure of the amplitude matrix to avoid over-smoothing of recovered speech. This could be a direction for future work.
  • Markov Chain Limitations: The paper acknowledges that the Markov chain model might not perfectly capture the dynamics of child's voice due to less fluent pronunciation.
  • Future Work: The authors suggest that their method can be applied to speech enhancement, which they demonstrate, and that further improvements might be possible by exploring amplitude structure.

Summary updated

BibTeX
@inproceedings{jia2020speech,
  author = {Xiaoli Jia and Peilin Liu and Sumxin Jiang},
  title = {{Speech Compressive Sampling Using Approximate Message Passing and a Markov Chain Prior}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=cOZUfk29_-o},
}
Figure