Robust statistical voice activity detection using a likelihood ratio sign test

Shiwen Deng, Jiqing HanInterspeech 2010

TL;DR

This paper introduces a novel Voice Activity Detection (VAD) method that leverages nonparametric detection theory by applying a sign test to the likelihood ratio (LR). This approach bypasses the need for exact statistical models of speech and noise, offering improved robustness in noisy environments compared to traditional model-based VADs.

What Problem Does This Paper Address?

Voice Activity Detection (VAD) is crucial for many speech processing applications, acting as a gatekeeper to distinguish speech from silence or background noise. Traditional VAD methods often rely on simple features like energy or zero-crossing rates, which perform poorly in noisy conditions. More advanced methods use statistical models and likelihood ratio tests (LRTs) to classify speech and nonspeech segments. However, these statistical models often struggle to accurately capture the complex and non-stationary nature of real-world noisy speech and various noise types. This mismatch between assumed models and actual data leads to decreased robustness and accuracy. The paper aims to address this by developing a VAD that is less dependent on precise statistical modeling.

Methods

The core innovation of this paper is the Likelihood Ratio Sign Test (LRST). Instead of directly comparing the likelihood ratio (LR) of speech versus nonspeech against a fixed threshold (as in conventional LRT-based VADs), the LRST treats the LR itself as a random variable.

Here's how it works:

  1. Multiple Observations: The algorithm processes multiple observation vectors (e.g., spectral features from short audio frames) within a sliding window.
  2. Log Likelihood Ratio (LLR): For each observation vector, a log likelihood ratio (LLR) is calculated.
  3. Sign Mapping: Each LLR is then mapped to a "sign" (1 if the LLR exceeds a predefined LRT threshold, 0 otherwise). This step is where the nonparametric aspect comes in – it doesn't require knowing the exact distribution of the LLR, only whether it's above or below a certain point.
  4. Sign Test: A decision is made based on the sequence of these signs over the sliding window. This is analogous to a sign test in statistics, which is robust to the exact distribution of the underlying data.
  5. Optimal Threshold Derivation: The paper also derives an optimal threshold for the LRST based on minimizing the Bayes error probability, considering prior probabilities of speech and nonspeech. They also show that approximating these prior probabilities as equal (0.5) yields a suitable optimal threshold.

For comparison, the paper uses the Multiple Observation Likelihood Ratio Test (MO-LRT) as a benchmark, which is a more sophisticated parametric approach.

Datasets

  • Training Data: A clean speech material of 125 seconds was used, manually labeled for active and inactive speech segments. This data was then corrupted with white, babble, and vehicle noise from the NOISEX-92 database at 5 dB SNR to create noisy training conditions.
  • Testing Data: A different 130-second speech segment from the TIMIT database was used for testing. Similar noise conditions (white, babble, vehicle) from NOISEX-92 were applied at various SNRs (5, 10, 15 dB).

Key Insights and Ideas

  • Robustness through Nonparametricity: The most significant insight is that by using the sign of the likelihood ratio, the VAD becomes less sensitive to the exact statistical distributions of speech and noise. This is crucial because real-world audio is rarely perfectly modeled by simple parametric distributions.
  • Sign Test for VAD: Applying the statistical sign test to the likelihood ratio is a novel way to achieve robust VAD. It leverages the inherent robustness of the sign test to distributional assumptions.
  • Optimal Threshold Simplification: The paper demonstrates that the optimal LRST threshold is relatively insensitive to the prior probabilities of speech and nonspeech, allowing for a practical approximation (e.g., assuming equal priors) that simplifies implementation.
  • Performance Gains in Noise: The LRST consistently outperforms the more complex MO-LRT in noisy conditions, highlighting the advantage of its nonparametric approach.

Key Results

The experimental results, visualized through Receiver Operating Characteristic (ROC) curves and summarized in Table 2, show that the proposed LRST VAD achieves better performance than the MO-LRT VAD across different noise types (Gaussian, babble, vehicle) and SNRs (5, 10, 15 dB).

For instance, at 5 dB SNR in Gaussian noise, the LRST achieves a detection probability ($P_d$) of 0.92 with a false alarm probability ($P_f$) of 0.08, compared to MO-LRT's $P_d$ of 0.78 and $P_f$ of 0.10. Similar improvements are observed across other noise conditions and SNRs, indicating a significant enhancement in detection accuracy and robustness.

Limitations and Future Work

  • Hang-over Scheme: The paper explicitly states that a "hang-over scheme" (a common VAD post-processing technique to smooth decisions) was not used in their experiments to focus solely on the statistical model and decision rule. This implies that incorporating such schemes could further improve performance.
  • Window Length: The optimal window length for the LRST was set to 13 frames, based on previous work. The paper notes that this value was considered constant across different noise environments, and exploring adaptive window length selection could be a future direction.
  • Parameter Analysis: While the paper analyzes the optimal threshold derivation and the effect of prior probabilities, a deeper dive into the sensitivity of other parameters (like the LRT threshold $\eta$) to different noise types and SNRs could be beneficial.
  • Real-time Implementation: The paper mentions that the m-frame delay imposed by the sliding window is not a serious implementation obstacle, but further work on optimizing the algorithm for real-time processing in diverse applications would be valuable.

Summary updated

BibTeX
@inproceedings{deng2010robust,
  author = {Shiwen Deng and Jiqing Han},
  title = {{Robust statistical voice activity detection using a likelihood ratio sign test}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure