Selective gammatone filterbank feature for robust sound event recognition

Yi Ren Leng, Tran Huy Dat, Norihide Kitaoka, Haizhou Li

Abstract

Automatic Sound Event Recognition (SER) often fails in real-world noisy environments due to the limitations of standard feature extraction methods like Mel-Frequency Cepstral Coefficients (MFCCs). This paper proposes a Selective Gammatone Filterbank (Sel. G) feature that avoids logarithmic compression and employs a statistical t-test to select the most noise-robust frequency channels. Experimental results demonstrate that this approach significantly improves recognition accuracy, achieving up to 81.03% accuracy in 0dB SNR car noise environments, compared to 40.08% for MFCCs.

Introduction

Sound Event Recognition (SER) is critical for applications ranging from smart home automation to advanced hearing aids. However, background noise remains a significant hurdle. Traditional systems rely on MFCCs, which were optimized for speech and utilize logarithmic dynamic range compression. While effective for clean signals, this compression can obscure the distinction between signal and noise in high-interference environments. As shown in Figure 1, at 0dB SNR, noise significantly pollutes various frequency channels, making event detection challenging.

Figure 1: Gammatone filterbank output at 0dB SNR, illustrating how noise pollutes frequency channels across time.

Methodology

The proposed Selective Gammatone Filterbank pipeline mimics the human ear's frequency selectivity while introducing two key innovations for robustness:

  1. Raw Dynamic Range: Instead of applying logarithmic compression, the system uses the raw output of the Gammatone filters. This preserves the intensity peaks of sound events, which serve as vital signatures against background noise.
  2. Statistical Channel Selection: A 36-channel filterbank is initially used. To mitigate the impact of noise, a t-test is applied to identify channels with the least variance between clean and noisy training samples. Only the 12 most robust channels are selected for the final feature vector.

These features, along with their delta components, are processed using a Hidden Markov Model (HMM) for classification.

Experimental Setup

The system was evaluated using the RWCP Sound Scene Database for core events (e.g., glass breaking), with noise profiles from CENSREC-4 and NOISEX-92 (e.g., car noise, canteen, factory). Models were trained on clean data (40dB SNR) and tested across a range of noise levels down to 0dB SNR.

Results and Discussion

The Selective Gammatone (Sel. G) feature consistently outperformed MFCCs and other variants. In the "Car" noise environment (Table 1), Sel. G maintained high accuracy even as noise levels increased.

Table 1: Recognition accuracy (%) in the Car noise environment across different SNR levels.

At 0dB SNR, Sel. G achieved 81.03% accuracy, a 100% relative improvement over the 40.08% achieved by MFCCs. Similar trends were observed in the "Canteen" environment, where Sel. G reached 73.73% compared to the 21.75% of MFCCs. These results highlight the effectiveness of discarding "noisy" data channels and maintaining raw dynamic range for machine-based recognition.

Conclusion

By prioritizing raw signal intensity and intelligently selecting frequency channels, the Selective Gammatone feature offers a robust alternative to speech-centric features for general sound event recognition. While there is a slight trade-off in perfectly clean conditions, the gains in noisy real-world environments are substantial. Future work will explore real-time dynamic channel adaptation to further enhance performance in changing acoustic landscapes.

Summary updated

BibTeX
@inproceedings{leng2010selective,
  author = {Yi Ren Leng and Tran Huy Dat and Norihide Kitaoka and Haizhou Li},
  title = {{Selective gammatone filterbank feature for robust sound event recognition}},
  booktitle = {Interspeech},
  year = {2010},
  doi = {10.21437/interspeech.2010-617},
}
Figure