A low-complexity voice activity detector for smart hearing protection of hyperacusic persons

Narimene Lezzoum, Ghyslain Gagnon, Jérémie VoixInterspeech 2013

TL;DR

This paper introduces a novel, low-complexity Voice Activity Detector (VAD) designed for smart hearing protection devices aimed at individuals with hyperacusis. By analyzing the temporal envelope of speech across different frequency bands using the Inter-Quartile Range (IQR) and a genetic algorithm for optimization, the VAD effectively distinguishes speech from noise, allowing speech to pass through while blocking unwanted sounds.

What Problem Does This Paper Address?

Hyperacusis, an extreme sensitivity to everyday sounds, significantly impacts individuals' quality of life, often leading to social isolation. While passive hearing protection devices (HPDs) offer some relief, they indiscriminately block all sounds, including desired speech. The development of "smart" HPDs that can selectively allow speech to pass through while attenuating noise is crucial. A key component of such smart HPDs is a Voice Activity Detector (VAD) that can accurately identify speech in real-time, even in noisy environments. Existing VADs often struggle with performance degradation in low Signal-to-Noise Ratios (SNRs) or require significant computational resources, making them unsuitable for embedded systems like HPDs. This paper addresses the need for a computationally efficient and robust VAD specifically tailored for this application.

Methods

The proposed VAD employs a multi-stage approach:

  • Windowing: The incoming audio signal is segmented into frames using a Hamming window with 80% overlap.
  • Filter Bank Analysis: Each frame is passed through a 16-band filter bank, implemented using 4th-order Butterworth filters. The cut-off frequencies are defined on the Bark scale, ranging from 20 Hz to 3150 Hz.
  • Temporal Envelope Extraction: For each frequency band, the temporal envelope is extracted using the Hilbert Transform.
  • Statistical Assessment of Temporal Envelopes: The Inter-Quartile Range (IQR) is calculated for the temporal envelopes within each frequency band. The IQR quantifies the dispersion of envelope magnitudes, serving as a feature to differentiate speech from noise.
  • Area Under IQR Curve: Recognizing that the first formant of speech is more robust to noise than higher formants, the VAD calculates an "Area under IQR Curve" (A1) specifically within the frequency bands that capture this crucial speech information (bands 3-9).
  • Normalization and Decision Ratios: To overcome the dependence of A1 on the absolute scale of the temporal envelope, two additional areas (A2 and A3) are calculated from the IQR in specific low and high frequency bands, respectively. These areas are used to normalize A1, creating ratios R1 and R2. These ratios are designed to be less sensitive to overall signal level variations.
  • Decision and Hangover Scheme: The VAD makes a decision based on two thresholds (T1 and T2) applied to R1 and R2. A "hangover" scheme is incorporated to prevent premature VAD deactivation during short speech pauses or unvoiced sounds. This scheme ensures that the VAD remains active for a specified duration after speech is detected.
  • Parameter Optimization: The decision thresholds (T1, T2), the hangover duration, and the number of consecutive frames required for speech confirmation (N) are optimized off-line using a Genetic Algorithm (GA). The objective function for the GA is the F1 score, which balances precision and recall, aiming to maximize speech detection accuracy while minimizing false positives and negatives.

Datasets

  • Speech Signals: TIMIT database.
  • Noise Environments: Subway, Exhibition, Babble, Airport, and Car noises from the AURORA database.
  • Training/Optimization Data: A limited subset of five speech signals corrupted by "Subway" noise at 0 dB SNR.
  • Validation Data: 90 speech signals corrupted by five different noise environments at three SNRs (10 dB, 5 dB, and 0 dB).

Key Insights and Ideas

  • IQR as a Robust Feature: The Inter-Quartile Range proves to be an effective statistical measure for characterizing the temporal envelope's dispersion, offering a robust way to distinguish speech from noise, especially in varying SNR conditions.
  • Focus on First Formant: The insight that the first formant is more resilient to noise than higher formants is leveraged by focusing the primary decision feature (A1) on the frequency bands where this formant is prominent.
  • Normalization for Robustness: Normalizing the primary speech feature (A1) with areas derived from noise-dominant frequency bands (A2, A3) creates ratios (R1, R2) that are less susceptible to overall signal level fluctuations, leading to more reliable decision-making.
  • GA for Optimal Parameter Tuning: Employing a Genetic Algorithm is a powerful approach to fine-tune multiple interdependent parameters (thresholds, hangover, observation count) simultaneously, ensuring optimal performance for the specific application.
  • Hangover Scheme for Intelligibility: The hangover mechanism is crucial for maintaining speech intelligibility by preventing the VAD from prematurely switching to noise detection during natural speech pauses or unvoiced segments.

Key Results

  • Improved F1 Score: The proposed VAD achieved an average F1 score of 85.9% across all tested SNRs and noise environments, significantly outperforming Sohn's VAD (77.5%).
  • Reduced Mid-Speech Clipping (MSC): The proposed VAD reduced the Mid-Speech Clipping rate by a factor of three compared to Sohn's VAD, indicating better preservation of speech intelligibility.
  • Computational Efficiency: The computational cost of the proposed VAD's unique steps (IQR and area calculations) is estimated to be around 18.4% of the total available processing power on a typical low-power DSP, making it feasible for embedded implementation.
  • Robustness to Long Durations: Validation with a 150-second speech signal confirmed that the VAD's performance is not solely dependent on the hangover duration but is inherently robust.

Limitations and Future Work

  • Limitations:
    • The optimization was performed using a limited dataset (five speech signals with "Subway" noise at 0 dB SNR).
  • The paper acknowledges that the hangover duration might seem long compared to telecommunications applications, but argues it's necessary for preserving entire speech signals in this context.
  • Future Work:
    • Validate the proposed VAD with subjective listening tests.
  • Develop noise reduction techniques to further enhance speech intelligibility.
  • Adapt the dynamic range of incoming speech signals to ensure they are transmitted to the protected ear without causing discomfort or damage.

Summary updated

BibTeX
@inproceedings{lezzoum2013low,
  author = {Narimene Lezzoum and Ghyslain Gagnon and Jérémie Voix},
  title = {{A low-complexity voice activity detector for smart hearing protection of hyperacusic persons}},
  booktitle = {Interspeech},
  year = {2013},
}
Figure