Alternative Frequency Scale Cepstral Coefficient for Robust Sound Event Recognition

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

Abstract

Standard Mel-Frequency Cepstral Coefficients (MFCCs), while highly effective for speech recognition, often falter in general Sound Event Recognition (SER) tasks, particularly in noisy environments. This paper investigates the limitations of the biologically-inspired Mel scale and proposes alternative linear and exponential frequency scales. Our findings demonstrate that these alternative scales provide superior robustness against low-frequency environmental noise and better capture the broad spectral characteristics of non-speech sounds.

Introduction

Sound Event Recognition (SER) involves identifying diverse acoustic signals such as breaking glass, gunshots, or alarms. Traditionally, researchers have adopted MFCCs—the gold standard in Automatic Speech Recognition (ASR)—for these tasks. However, the Mel scale is specifically tuned to human speech perception, which prioritizes lower frequencies. This paper argues that this speech-centric approach is suboptimal for general sound events and proposes a shift toward more flexible frequency scaling.

The Case Against the Mel Scale

The authors identify two primary issues with using MFCCs for general sound recognition: spectral mismatch and noise sensitivity.

Spectral Mismatch

Unlike speech, which is predominantly low-frequency, many general sound events exhibit significant energy across a much wider spectral range. The Mel scale's logarithmic compression "zooms in" on low frequencies, potentially discarding discriminative high-frequency information essential for identifying sounds like bells or metallic strikes.

Spectral characteristics of different sound classes: (a) Bells, (b) China, (c) Speech, and (d) Wood. Note the broad spectral spread of non-speech sounds compared to the low-frequency concentration of speech.

Noise Sensitivity

Most common environmental noises (e.g., car engines, air conditioning, background babble) are concentrated in the low-frequency region. Because the Mel scale allocates a higher density of filters to these frequencies, it inadvertently captures more noise. This leads to a significant performance degradation when there is a mismatch between training (clean) and testing (noisy) conditions.

Proposed Alternative Scales

To address these flaws, the researchers modified the standard MFCC extraction pipeline by replacing the Mel filterbank with two alternatives:

  1. Linear Scale: Filters are distributed uniformly across the entire frequency range, providing equal weight to high-frequency components.
  2. Exponential Scale: This scale effectively inverts the Mel approach, allocating more filters to higher frequencies and fewer to the noise-prone low-frequency region.

These features were then classified using a Proximal Support Vector Machine (PSVM).

Experimental Evaluation

The study utilized a diverse dataset including 19 classes from the RWCP Sound Scene Database and voiced sounds like laughter and screams. Signals were corrupted with various background noises (Canteen, Car, Hall, Office) at Signal-to-Noise Ratios (SNRs) ranging from 15dB to 0dB.

Results and Discussion

The experimental results strongly favored the proposed scales in noisy environments:

  • Exponential Superiority: For general sounds, the Exponential scale was the top performer. At 15dB SNR, it achieved a 60% relative improvement over MFCCs. In extreme noise (0dB), the improvement reached nearly 100%.
  • Robustness: The Exponential scale acts as a natural high-pass emphasis, effectively ignoring the "noise floor." Euclidean distance measurements confirmed that Exponential features remain more stable than MFCCs as noise levels increase.
  • Voiced Sound Performance: Surprisingly, even for speech-like sounds, the Exponential scale remained robust, outperforming MFCCs in most noisy scenarios, albeit with a smaller margin than for generic sounds.

Conclusion

This research demonstrates that the "one-size-fits-all" application of speech-centric features is a liability for robust sound event recognition. By adopting Linear or Exponential frequency scales, systems can achieve significantly higher accuracy in real-world, noisy environments without the need for complex denoising algorithms. Future work should explore task-specific optimized scales for specialized applications like surveillance or industrial monitoring.

Summary updated

BibTeX
@inproceedings{leng2011alternative,
  author = {Yi Ren Leng and Tran Huy Dat and Norihide Kitaoka and Haizhou Li},
  title = {{Alternative Frequency Scale Cepstral Coefficient for Robust Sound Event Recognition}},
  booktitle = {Interspeech},
  year = {2011},
  doi = {10.21437/interspeech.2011-114},
}
Figure