Power Spectral Deviation-Based Voice Activity Detection Incorporating Teager Energy for Speech Enhancement

Sang-Kyun Kim, Sang-Ick Kang, Young-Jin Park, Sanghyuk Lee, Sangmin LeeSymmetry 2016 · DOI

TL;DR

This paper introduces a novel Voice Activity Detection (VAD) algorithm that significantly improves speech/non-speech discrimination in noisy environments. By incorporating Teager Energy (TE) into Power Spectral Deviation (PSD) calculations and employing a TE-based likelihood ratio and speech absence probability, the proposed method achieves more robust and accurate VAD performance compared to existing techniques.

What Problem Does This Paper Address?

Accurate Voice Activity Detection (VAD) is a critical component for many speech processing tasks, including speech recognition and noise suppression. However, existing VAD algorithms often struggle in noisy environments, leading to errors in identifying speech segments. Traditional VAD features like short-time energy or zero-crossing rate are highly sensitive to noise. While methods like Power Spectral Deviation (PSD) offer some robustness, they can still be insufficient. The challenge is to develop a VAD system that can reliably distinguish speech from various types of noise, even at low signal-to-noise ratios (SNRs).

Methods

The core of the proposed VAD algorithm lies in its novel feature extraction and decision-making process:

  • Teager Energy (TE) Operator: The paper leverages the Teager Energy (TE) operator, known for its ability to enhance the discriminability between speech and noise. TE is applied to the noisy speech signal.
  • TE-based Power Spectral Deviation (TE-PSD): Instead of using conventional PSD, the authors propose a TE-based PSD (TE-PSD) as the primary feature for VAD. This TE-PSD is calculated using the TE of the noisy speech signal.
  • TE-based Likelihood Ratio (LR) and Speech Absence Probability (SAP): To further refine the VAD performance, the algorithm incorporates a TE-based Likelihood Ratio (LR) and Speech Absence Probability (SAP). These are derived from the TE of the speech and noise signals and are used to modify the TE-PSD.
  • Smoothing and Decision Rule: The TE-PSD is smoothed using the SAP, and a final decision is made based on a threshold applied to the modified TE-PSD. The algorithm also includes a noise power estimation update mechanism.

Datasets

The paper does not explicitly mention training datasets. For evaluation, they used a dataset of 456 seconds of speech data recorded by four males and four females, sampled at 8 kHz. This data was corrupted with three types of noise (babble, car, and office noises) at various SNRs (0, 5, 10, and 15 dB). Reference decisions were made by manually labeling clean speech frames.

Key Insights and Ideas

  • TE Enhances Discriminability: The Teager Energy operator proves effective in amplifying speech signal characteristics while suppressing noise, leading to better feature representation for VAD.
  • TE-PSD Outperforms Conventional PSD: By incorporating TE into PSD calculations, the proposed TE-PSD feature offers superior discrimination between speech and noise compared to traditional PSD, especially in challenging noisy conditions.
  • LR and SAP Refine Decision Making: The integration of TE-based Likelihood Ratio and Speech Absence Probability acts as a sophisticated weighting and smoothing mechanism, further improving the robustness of the VAD decision.
  • Robustness Across Noise Types and SNRs: The proposed method demonstrates consistent performance improvements across different noise environments and a wide range of SNRs, highlighting its practical applicability.

Key Results

The proposed TE-based VAD algorithm demonstrated superior performance compared to conventional methods, including CMAP-SG and ITU-T G.729B Appendix III.

  • Objective Evaluation: Comparative results showed lower Total Error Rates (TER), False Rejection Rates (FRR), and False Acceptance Rates (FAR) for the proposed method across various noise conditions and SNRs. For instance, at 0 dB SNR in babble noise, the proposed method achieved a TER of 19.52%, compared to 21.55% for G.729B App. III and 22.76% for CMAP-SG.
  • Receiver Operating Characteristic (ROC) Curves: ROC curves consistently showed that the proposed method achieved higher detection rates (100-FRR) for a given false acceptance rate (FAR) compared to the other methods, particularly at lower SNRs.
  • Speech Quality (PESQ): Objective speech quality evaluation using PESQ scores indicated that the proposed VAD, when integrated into a noise suppression system, yielded comparable or slightly better speech quality than conventional VAD methods, especially at low SNRs.

Limitations and Future Work

  • Computational Complexity: While not explicitly detailed as a limitation, the use of TE and other complex calculations might imply a higher computational cost compared to simpler VAD methods.
  • Parameter Tuning: Although the paper mentions that ROC curves are insensitive to parameter tuning, the selection of optimal parameters (like q in Equation 12) was based on experimental results. Further research could explore adaptive parameter selection.
  • Future Work: The authors suggest that the proposed TE-based VAD algorithm can be further improved by exploring more advanced noise estimation techniques and potentially integrating it with other speech enhancement or recognition modules for end-to-end system optimization.

Summary updated

BibTeX
@article{kim2016power,
  author = {Sang-Kyun Kim and Sang-Ick Kang and Young-Jin Park and Sanghyuk Lee and Sangmin Lee},
  title = {{Power Spectral Deviation-Based Voice Activity Detection Incorporating Teager Energy for Speech Enhancement}},
  journal = {Symmetry},
  year = {2016},
  doi = {10.3390/sym8070058},
}
Figure