Acoustic Features for Classification Based Speech Separation

Yuxuan Wang, Kun Han, DeLiang WangInterspeech 2012 · DOI

Abstract

Researchers at The Ohio State University have significantly improved monaural speech separation by expanding the acoustic feature set used to distinguish speech from noise. By employing Group Lasso for feature selection, they identified a powerful combination of speech recognition features—specifically GFCCs, MFCCs, and RASTA-PLP—paired with pitch. This "COMP+PITCH" set outperforms existing methods, particularly in generalizing to unseen background noises.

Introduction

Monaural speech separation, often referred to as the "cocktail party problem," involves isolating a target speaker from a background of interfering sounds using only a single microphone. A robust approach to this problem is to frame it as a binary classification task. The objective is to estimate an Ideal Binary Mask (IBM), which labels time-frequency (T-F) units as either speech-dominant ('1') or noise-dominant ('0').

Historically, this classification has relied on a limited set of features, primarily pitch and Amplitude Modulation Spectrograms (AMS). While effective for voiced speech, pitch fails for unvoiced segments, and AMS often lacks robustness across varying noise environments. This research explores whether features traditionally used in speech and speaker recognition can enhance separation performance.

Methodology

The proposed system utilizes a 64-channel gammatone filterbank to process audio, followed by a bank of Support Vector Machines (SVMs) trained for each frequency subband.

Feature Exploration and Selection

The study evaluated several features common in speech recognition:

  • GFCC (Gammatone Frequency Cepstral Coefficients)
  • MFCC (Mel Frequency Cepstral Coefficients)
  • PLP (Perceptual Linear Prediction)
  • RASTA-PLP (Relative Spectral Transform - PLP)

To manage the high dimensionality of these features and prevent overfitting, the researchers applied Group Lasso. This statistical method treats each feature type as a single group, selecting only those that significantly contribute to classification accuracy. This process led to the identification of the "COMP" set: a complementary combination of AMS, RASTA-PLP, and MFCC.

Experimental Setup

The system was trained on 50 IEEE utterances from a female speaker mixed with three types of noise: bird chirps, crow noise, and cocktail party noise (matched conditions). Testing was conducted on 20 new utterances mixed with both the training noises and three "unseen" noises: playground noise, traffic noise, and electric fan noise (unmatched conditions). All mixtures were set at a 0 dB Signal-to-Noise Ratio (SNR).

Results and Discussion

Performance was measured using the HIT−FA metric (the difference between the percentage of correctly identified speech units and incorrectly identified noise units) and SNR gain.

Matched vs. Unmatched Performance

In matched noise conditions, the COMP+PITCH system achieved a HIT-FA of 75%, surpassing the baseline of previous methods.

Table showing performance of various feature combinations under matched noise conditions.

The advantage of the proposed feature set became even more apparent in unmatched conditions. While previous methods saw significant performance drops, the COMP+PITCH system maintained a robust 68% HIT-FA.

Table showing performance of various feature combinations under unmatched (unseen) noise conditions.

Key Insights

  • GFCC Effectiveness: GFCCs proved highly effective in matched conditions, likely due to their ability to capture specific speaker characteristics.
  • RASTA-PLP Robustness: RASTA-PLP demonstrated remarkable stability across new noises by filtering out slow spectral changes and focusing on speech-characteristic modulations.
  • The Role of Pitch: As an intrinsic property of speech, pitch remains a reliable anchor even when other features degrade in high-noise environments. Combining the "COMP" feature set with pitch provided the most robust results across both matched and unmatched conditions.
  • Feature Complementarity: The results confirm that combining features with different strengths (e.g., AMS and RASTA-PLP) provides a more comprehensive representation for the classifier.

Conclusion

By integrating a broader range of acoustic features and utilizing systematic feature selection, this research demonstrates a significant leap in monaural speech separation performance, particularly for unseen noise environments. Future work will explore the generalization of these features across different speakers and genders, as well as the potential for real-time implementation and integration into multi-microphone systems.

Summary updated

BibTeX
@inproceedings{wang2012acoustic,
  author = {Yuxuan Wang and Kun Han and DeLiang Wang},
  title = {{Acoustic Features for Classification Based Speech Separation}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-432},
}
Figure