Noise robust voice activity detection using features extracted from the time-domain autocorrelation function

Houman Ghaemmaghami, Brendan Baker, Robert Vogt, Sridha SridharanInterspeech 2010

TL;DR

This paper introduces a novel Voice Activity Detection (VAD) method called AZR, specifically designed for noisy environments. By fusing two new features derived from the autocorrelation function – one measuring the maximum peak (MaxPeak) and another assessing signal periodicity (CrossCorr) – AZR significantly outperforms existing VAD systems in challenging, high-noise conditions.

What Problem Does This Paper Address?

Voice Activity Detection (VAD) is a crucial component in many speech processing applications, from automatic speech recognition to speaker diarization. The core challenge lies in accurately distinguishing between speech and non-speech segments, especially in the presence of significant background noise. Traditional VAD methods often struggle when the Signal-to-Noise Ratio (SNR) drops below 5 dB. This paper tackles this problem by developing a VAD system that is robust to high noise levels, focusing specifically on the reliable detection of voiced speech segments, which are inherently more structured and less random than unvoiced speech or noise.

Methods

The proposed AZR VAD system is a fusion of two novel features extracted from the time-domain autocorrelation function:

  • MaxPeak: This feature utilizes the maximum peak value of the normalized autocorrelation function within the lag range corresponding to expected human speech pitch periods (2 to 20 ms). While a standard approach, its robustness decreases with increasing noise.
  • CrossCorr: This is a new feature designed to be more noise-robust. It leverages the quasi-periodic nature of voiced speech. The method first identifies potential "periods" within the autocorrelation function by counting zero-crossing points. Then, it cross-correlates these segments to measure their similarity, providing an indicator of signal periodicity. This feature is hypothesized to be more resilient to noise than MaxPeak because the periodicity of voiced speech can be preserved even when the peak amplitude diminishes in noisy conditions.

These two features are fused using a weighted sum, with equal weights found to be optimal. To complete the VAD process, unvoiced speech segments are estimated using a smoothing filter, acknowledging that voiced speech is typically surrounded by unvoiced speech.

Datasets

The primary dataset used for evaluation is the QUT-NOISE-TIMIT corpus. This corpus consists of 600 hours of noisy speech, created by mixing clean speech from the TIMIT database with realistic noise recordings from the QUT-NOISE corpus. The dataset features 5 distinct real-life noise scenarios (CAFE, CAR, HOME, REVERB, STREET), each divided into two locations. Noise levels are varied across low (10-15 dB SNR), medium (0-5 dB SNR), and high (-5 to -10 dB SNR) settings. A total of 24,000 files were generated for testing.

Key Insights and Ideas

  • Periodicity as a Noise Robust Feature: The core insight is that the periodicity of voiced speech, as captured by the CrossCorr feature, is more robust to noise than simply looking at the maximum peak of the autocorrelation function. This is because the underlying periodic structure can persist even when the signal's amplitude is significantly degraded.
  • Fusion for Enhanced Robustness: Combining the MaxPeak and CrossCorr features through weighted sum fusion leads to a synergistic effect, yielding a VAD system that is more accurate than either feature used alone, particularly in high-noise scenarios.
  • Focus on Voiced Speech Detection: By explicitly targeting voiced speech, the AZR system simplifies the problem and leverages the inherent structure of voiced sounds, which are more distinguishable from noise than unvoiced sounds.
  • Importance of Realistic Noise: The use of the QUT-NOISE-TIMIT corpus, featuring real-world noise recordings, is crucial for evaluating VAD performance in practical, challenging conditions.

Key Results

The AZR VAD system demonstrated superior performance compared to several state-of-the-art and standardized VAD methods, including LTSD, Sohn's LRT, ITU-T G.729 Annex B, and ETSI AFE. Across all tested noise levels and scenarios, AZR achieved an average relative improvement of 24.8% in Half-Total Error Rate (HTER) compared to the best performing baseline system (LTSD). Specifically, AZR showed relative improvements of 33.6%, 28.5%, and 12.2% in HTER for low, medium, and high-noise scenarios, respectively. The system also showed better resilience to reverb compared to some baseline methods.

Limitations and Future Work

  • Focus on Voiced Speech: While the focus on voiced speech detection is a strength, the method relies on smoothing to estimate unvoiced segments. The accuracy of this estimation could be a limitation.
  • Computational Cost: The paper mentions that a non-overlapping frame implementation was chosen to reduce computation time. Further optimization for real-time applications might be necessary.
  • Future Work: The authors suggest that further exploration of feature fusion techniques and potentially incorporating other noise-robust features could lead to even greater improvements. The effectiveness of the AZR system in different acoustic environments and with various types of noise could also be investigated further.

Summary updated

BibTeX
@inproceedings{ghaemmaghami2010noise,
  author = {Houman Ghaemmaghami and Brendan Baker and Robert Vogt and Sridha Sridharan},
  title = {{Noise robust voice activity detection using features extracted from the time-domain autocorrelation function}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure