Toward detecting voice activity employing soft decision in second-order conditional MAP

Sang-Kyun Kim, Jae-Hun Choi, Sang-Ick Kang, Ji-Hyun Song, Joon-Hyuk ChangInterspeech 2010 · DOI

Introduction

Voice Activity Detection (VAD) serves as a fundamental component in modern speech processing systems, acting as a gatekeeper that distinguishes human speech from background noise. In mobile Voice over IP (VoIP) applications, an effective VAD is crucial for conserving bandwidth and extending battery life by suppressing transmission during silent periods. However, the unpredictable nature of background noise—ranging from street traffic to office chatter—poses a significant challenge for traditional statistical VAD models.

The Limitation of First-Order Models

Most conventional statistical VADs operate on a frame-by-frame basis or utilize first-order correlations, where the decision for the current frame depends only on the immediate predecessor. Human speech, however, possesses a natural temporal structure that extends beyond the typical 10–30ms frame. By failing to account for longer-term correlations, these systems often suffer from increased "misses" (misclassifying speech as noise) or "false alarms" (misclassifying noise as speech) during transitions.

Proposed Methodology: Second-Order CMAP with Soft Decisions

To address these limitations, this research introduces a Second-order Conditional Maximum A Posteriori (CMAP) approach combined with a soft decision scheme.

1. Second-Order Modeling

Unlike previous methods that only consider the status of the single previous frame, the proposed model looks back two frames. This results in four distinct historical states: Silence-Silence, Silence-Speech, Speech-Silence, and Speech-Speech. Each state is assigned its own optimized decision threshold, allowing the system to adapt its sensitivity based on the recent "trajectory" of the audio signal.

2. Soft Decision Scheme

A common pitfall in noisy environments is the propagation of errors; a "hard" incorrect decision in a previous frame can lead to an inappropriate threshold for the current one. The authors mitigate this by employing a Soft Decision scheme. Instead of switching abruptly between thresholds, the system calculates the Speech Absence Probability (SAP) for previous frames. The final threshold is a weighted average of the four potential thresholds, where weights are determined by the probability of being in each historical state.

Experimental Evaluation

The algorithm was rigorously tested against 456 seconds of speech data across four noise environments: Car, Street, Office, and Babble. Performance was evaluated at Signal-to-Noise Ratios (SNR) of 0, 5, 10, and 15 dB. Benchmarks included the G.729 Annex B standard, the Sohn et al. statistical model, and the first-order CMAP method.

Key Results

The second-order CMAP with soft decisions consistently outperformed all benchmarks across various metrics.

ROC curve showing the superior detection probability of the proposed soft-decision method compared to standard and first-order models.

As shown in the Receiver Operating Characteristic (ROC) curves, the proposed method achieves a higher detection probability for any given false alarm rate. Quantitatively, the method demonstrated a significant reduction in the overall probability of error ($P_E$) and the miss rate ($P_M$).

Comparative performance table showing Probability of Error (PE), Miss Rate (PM), and False Alarm Rate (PF) across different noise environments and SNRs.

For instance, in "Office" noise at 0dB SNR, the miss rate was reduced to 10.71%, a notable improvement over the 13.45% recorded for the Sohn et al. method. In "Street" noise at 5dB SNR, the total error rate dropped from 10.29% (first-order) to 10.15% with the proposed second-order soft-decision approach.

Conclusion

By incorporating longer-term temporal context and a robust soft-decision mechanism, the proposed VAD algorithm significantly enhances speech detection accuracy in challenging acoustic environments. While the second-order model increases computational complexity compared to first-order systems, the performance gains in noisy conditions justify its use for high-quality real-time communication.

Summary updated

BibTeX
@inproceedings{kim2010toward,
  author = {Sang-Kyun Kim and Jae-Hun Choi and Sang-Ick Kang and Ji-Hyun Song and Joon-Hyuk Chang},
  title = {{Toward detecting voice activity employing soft decision in second-order conditional MAP}},
  booktitle = {Interspeech},
  year = {2010},
  doi = {10.21437/interspeech.2010-767},
}
Figure