A framework for robust MFCC feature extraction using SNR-dependent compression of enhanced mel filter bank energies

Babak Nasersharif, Ahmad AkbariInterspeech 2006

TL;DR

This paper introduces a novel method to improve the performance of Mel-frequency Cepstral Coefficients (MFCCs) in noisy environments. By combining sub-band spectral subtraction with an innovative SNR-dependent compression of Mel filter bank energies, they achieve significant reductions in word error rates, particularly at low signal-to-noise ratios.

What Problem Does This Paper Address?

Traditional MFCCs, a cornerstone of speech recognition, are notoriously sensitive to background noise. This sensitivity leads to a sharp decline in recognition accuracy when speech is corrupted by additive noise. While various noise reduction techniques exist, they often struggle to effectively compensate for noise without introducing their own artifacts or failing to adapt to varying noise conditions. This paper aims to bridge this gap by developing a more robust feature extraction method that specifically targets and mitigates the impact of noise on MFCCs.

Methods

The proposed framework enhances MFCC feature extraction through a two-step process applied to Mel filter bank energies:

  1. Mel Sub-Band Spectral Subtraction: Instead of applying spectral subtraction to the entire speech spectrum, this method operates on individual Mel sub-bands. This localized approach aims to more precisely remove noise components within each frequency band. The process involves estimating noise in each sub-band and then subtracting it from the noisy speech energy, with parameters for over-estimation and spectral flooring adjusted per sub-band.
  2. SNR-Dependent Sub-Band Energy Compression: This is the core innovation. Conventional MFCCs use a logarithmic function to compress Mel filter bank energies. This paper replaces the logarithm with a novel compression function that is dependent on the Signal-to-Noise Ratio (SNR) within each Mel sub-band. The function uses a parameter w_i which adapts based on the sub-band SNR. When the SNR is high, w_i approaches a constant root (similar to root cepstral analysis), and when the SNR is low, w_i becomes smaller, effectively reducing the impact of noisy sub-bands. This SNR-dependent compression aims to better preserve information from cleaner sub-bands while downplaying contributions from noisier ones, offering a more nuanced approach than a fixed compression or a simple root function.

Datasets

The experiments were conducted using the TIMIT database for isolated word recognition. The authors selected specific sentences and speakers, creating a training set of 2349 utterances from 114 speakers and a testing set of 777 utterances from 37 speakers. Three types of additive noises—white, pink, and factory noises—were added to both training and testing sets, sourced from the NOISEX92 database.

Key Insights and Ideas

  • Sub-band processing is key: Operating spectral subtraction at the Mel sub-band level, rather than the full spectrum, allows for more targeted noise removal.
  • SNR-dependent compression is superior to fixed compression: The adaptive nature of the proposed compression function, which adjusts based on the local SNR of each sub-band, proves more effective than fixed compression methods (like logarithm or constant root) in noisy conditions.
  • Noise impacts different sub-bands differently: The SNR-dependent compression implicitly acknowledges that noise affects various frequency bands with varying severity, allowing the system to dynamically adapt its feature extraction.
  • Logarithm's limitations in noise: The paper highlights that the standard logarithmic compression in MFCCs can amplify distortions that are minor in the power spectrum, making them more susceptible to noise. The proposed method aims to circumvent this.

Key Results

The proposed method, termed CMSBS (Compression and Mel Sub-Band Spectral subtraction), demonstrated significant improvements in word error rate (WER) across various noise types and SNRs.

  • At an SNR of 0 dB, CMSBS achieved word error rates of 10.15% (white noise), 7.75% (pink noise), and 6.84% (factory noise).
  • In comparison, conventional MFCCs had word error rates exceeding 80% at 0 dB SNR for all noise types.
  • The proposed method reduced word error rates by approximately 70% at 0 dB SNR compared to conventional MFCCs.
  • CMSBS consistently outperformed other evaluated methods, including LMSBS (Logarithm and Mel Sub-Band Spectral subtraction) and RMFCC (Root MFCC), across all tested conditions.

Limitations and Future Work

  • Limitations: The paper notes that while the proposed method significantly improves robustness, it still has limitations in very low SNR conditions. The effectiveness of spectral subtraction can be limited when noise levels are extremely high.
  • Future Work: The authors plan to further optimize the proposed compression function. Additionally, they suggest incorporating voice activity detectors (VADs) to improve the accuracy of noise estimation in Mel sub-bands, which could lead to even better performance.

Summary updated

BibTeX
@inproceedings{nasersharif2006framework,
  author = {Babak Nasersharif and Ahmad Akbari},
  title = {{A framework for robust MFCC feature extraction using SNR-dependent compression of enhanced mel filter bank energies}},
  booktitle = {Interspeech},
  year = {2006},
}
Figure