On the use of machine learning methods for speech and voicing classification

Philip Harding, Ben MilnerInterspeech 2012

TL;DR

This paper investigates the effectiveness of various machine learning (ML) classifiers for voice activity detection (VAD) and voicing classification (VC) in noisy environments. The authors found that MLPs, Rotation Forests, and Gaussian Mixture Models (GMMs) significantly outperform traditional methods and other ML classifiers, especially in challenging low-SNR conditions.

What Problem Does This Paper Address?

Voice Activity Detection (VAD) and Voicing Classification (VC) are fundamental tasks in speech processing, crucial for applications like speech recognition, coding, and noise estimation. Traditional methods often rely on hand-crafted features and thresholds, making them sensitive to noise. While machine learning has been applied to these problems, a comprehensive comparison of various ML classifiers under realistic noisy conditions was lacking. This paper aims to bridge that gap by evaluating a broad spectrum of ML techniques and comparing their robustness against conventional approaches.

Methods

The study employs Mel-Frequency Cepstral Coefficients (MFCCs) as the primary feature set, augmented with their temporal derivatives (velocity and acceleration) to capture dynamic speech characteristics. A variety of ML classifiers are evaluated, including:

  • Parametric/Probabilistic: Gaussian Mixture Models (GMMs) and Naïve Bayes.
  • Non-probabilistic: Support Vector Machines (SVMs) and Logistic Regression.
  • Artificial Neural Networks: Multilayer Perceptrons (MLPs).
  • Tree-based: C4.5 Decision Trees and Rotation Forests (an ensemble method).

The GMM classifier uses an in-house implementation, while others are based on the WEKA API. The paper also includes two industry-standard VADs, G.729 Annex B and ETSI XAFE, as baselines.

Datasets

The speech data used in this study is derived from the WSJ-CAM0 dataset. For evaluation, speech was contaminated with:

  • Stationary Noise: White noise.
  • Non-stationary Noise: Street noise from the NOIZEUS dataset.

The Modified Intermediate Reference System (IRS) filter was applied to simulate telephone channel characteristics. Classifiers were trained on 20 male and 20 female talkers and tested on 5 male and 5 female talkers previously unseen. Reference VAD data was generated using an energy threshold and pitch tracking, with subsequent manual correction.

Key Insights and Ideas

  • ML Dominance in Noise: Machine learning classifiers, particularly MLPs, Rotation Forests, and GMMs, demonstrate significantly superior robustness to noise compared to traditional VAD methods (G.729, ETSI XAFE).
  • Feature Dynamics Matter: Augmenting static MFCCs with temporal derivatives (velocity and acceleration) consistently improves performance across most classifiers, highlighting the importance of capturing speech dynamics.
  • Correlation Handling is Key: The top-performing classifiers (MLP, Rotation Forest, GMM) share an ability to handle correlated feature vectors. This is crucial because temporal derivatives reintroduce correlation, which can challenge simpler models.
  • Unvoiced Speech is Tricky: Voicing classification is most challenging for unvoiced speech, especially at low SNRs. This is attributed to the limited distinguishing features and the impact of bandwidth limitations and filtering on high-frequency energy cues.
  • Baseline Limitations: Conventional VADs like G.729 and ETSI XAFE, while performing reasonably well at high SNRs, degrade rapidly at lower SNRs, underscoring the need for more robust methods.

Key Results

  • VAD Performance: In ROC space, MLPs, Rotation Forests, and GMMs consistently outperformed other classifiers and the baseline VADs across various SNRs and noise types. Even at 0dB SNR in street noise, these ML methods achieved respectable performance, while baselines degraded significantly.
  • Voicing Classification Accuracy: Similar trends were observed for voicing classification. Rotation Forest and MLP achieved the highest overall accuracy. While ML methods struggled with unvoiced speech identification at low SNRs, they generally outperformed the ETSI XAFE baseline.
  • Feature Impact: Adding CO to C12 MFCCs with velocity and acceleration increased accuracy by 7-9% compared to static MFCCs.

Limitations and Future Work

The authors acknowledge that not all possible classifiers could be evaluated. They postulate that further improvements for some classifiers (like SVM and Logistic Regression) could be achieved by decorrelating the feature vector, for example, using Principal Component Analysis (PCA). This suggests a promising avenue for future research.

Summary updated

BibTeX
@inproceedings{harding2012use,
  author = {Philip Harding and Ben Milner},
  title = {{On the use of machine learning methods for speech and voicing classification}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure