Weighted MFCC (WMFCC) Techniques

• Weighted MFCC (WMFCC) is an advanced variant of classical MFCC that applies entropy-based or learnable weighting schemes to enhance audio feature representation.
• It improves feature discriminability by purposefully emphasizing high-order cepstral coefficients often ignored in standard MFCC pipelines.
• WMFCC systems integrate parameterized stages for windowing, FFT, filterbanks, and DCT to enable end-to-end learning while maintaining computational stability.

Weighted Mel-Frequency Cepstral Coefficients (WMFCC) are variants of the classical Mel-Frequency Cepstral Coefficients (MFCC) framework, intended to increase the representational capacity and adaptivity of cepstral features in audio processing pipelines for tasks such as speaker verification and pathological speech analysis. WMFCC architectures can involve explicit weighting schemes applied post-MFCC extraction—typically to enhance the discriminatory contribution of high-order coefficients—or the introduction of learnable parameters into each linear transform within the MFCC computation stack, enabling full differentiability and data-driven adaptation in deep learning systems (Xu et al., 2018, Liu et al., 2021).

1. Motivation and Distinction from Standard MFCC

The standard MFCC extraction pipeline comprises pre-emphasis, framing, windowing, FFT, Mel-scale filterbank application, logarithmic compression, and discrete cosine transform (DCT), followed by truncation to a fixed number of low-order coefficients. This process discards or underutilizes high-order cepstral coefficients, as their magnitudes are reduced to near zero—diminishing their utility for downstream classification. The rationale for WMFCC approaches is to rectify this imbalance. Two core methodologies exist:

• Applying a per-coefficient weight—either computed using an entropy-based measure or learned end-to-end—so that all cepstral dimensions contribute meaningfully to classification (Xu et al., 2018).
• Replacing each fixed linear operator in the MFCC pipeline (window, DFT, Mel filterbank, DCT) with a parameterized, learnable counterpart jointly optimized with the task-specific loss (Liu et al., 2021).

Both paradigms enable more flexible and informative representations of speech signals for automatic classification.

2. Entropy-Based WMFCC: Mathematical Formulation and Implementation

In the entropy-weighted MFCC formulation, the weighting vector $\mathbf{w} = (w_1, \dots, w_D)$ is derived from entropy calculations over the MFCC matrix for a given utterance. The steps are:

1. MFCC Stacking: For a sample segmented into $N$ frames, each with $D$ MFCCs, form $$M = [\,\mathbf{m}_1\;\mathbf{m}_2\;\cdots\;\mathbf{m}_N\,] \in \mathbb{R}^{D \times N}$$.
2. Row Normalization: Normalize each row (coefficient index $j$) of $M$ to $[0,1]$:

$$\tilde m_{ij} = \frac{m_{ij} - \min_k m_{kj}}{\max_k m_{kj} - \min_k m_{kj}}$$

1. Probability-like Assignment: For each entry, compute

$$y_{ij} = \frac{\tilde m_{ij}}{\sum_{i=1}^N \tilde m_{ij}}$$

2. Entropy Calculation: For each coefficient $j$,

$$e_j = -k \sum_{i=1}^N y_{ij} \ln(y_{ij}), \qquad k = \frac{1}{\ln N}$$

so that $0 \le e_j \le 1$.

3. Information Content Weighting:

$w_j = \frac{1 - e_j}{\sum_{p=1}^D (1 - e_p)}$

with normalization $\sum_j w_j = 1$.

4. Weight Application: Weighted MFCCs are then

$$\tilde c_j = w_j\,c_j, \qquad \tilde{\mathbf{c}} = \mathbf{w}\circ\mathbf{c}$$

where $\circ$ denotes element-wise product and $\mathbf{c}$ is the D-dimensional cepstral vector (Xu et al., 2018).

Weights are computed per sample (or per speaker) and applied after the DCT step, before any subsequent normalization or liftering procedures.

3. Learnable WMFCC: Differentiable Pipeline Components

A distinct WMFCC variant replaces the MFCC’s fixed linear steps with parameterized, learnable kernels (Liu et al., 2021):

• Window: Learnable window vector $w \in \mathbb{R}^M$ replaces the fixed Hamming window; for each frame $x$, $x_w = w \odot x$.
• DFT: Real and imaginary components $F_1, F_2 \in \mathbb{R}^{K \times M}$ are adapted, yielding power spectrum $P = (F_1 x_w)^2 + (F_2 x_w)^2$.
• Mel-filterbank: Adaptable matrix $M(\theta_m) \in \mathbb{R}^{L \times K}$, initialized from the standard Mel filterbank and regularized for non-negativity.
• DCT: Learnable DCT matrix $D(\theta_d) \in \mathbb{R}^{C \times L}$, typically regularized toward orthonormality.
• End-to-End Optimization: These parameters are jointly trained via backpropagation through the MFCC computation, facilitating end-to-end feature adaptation.

This approach retains interpretability while aligning feature extraction to the target classification objective.

4. Signal Processing Workflow and WMFCC Injection Point

Both methodologies align with the canonical MFCC workflow up to DCT:

Stage	Standard MFCC	WMFCC Variation
Pre-emphasis	Fixed ($k \sim 0.97$)	Same / or learned
Windowing	Fixed (Hamming)	Learnable vector
FFT (DFT)	Fixed (DFT matrix)	Learnable kernels
Mel-filterbank	Fixed triangular	Learnable matrix
Log compression	$\log(\cdot)$	Same
DCT	Fixed (DCT-II)	Learnable matrix
Weighting	—	Entropy-based / learned

For entropy-based WMFCC, weights are applied after DCT. In neural WMFCC, each pipeline stage is potentially learnable, with constraints ensuring stability (e.g., non-negativity, orthonormality).

5. Empirical Results and Application Domains

WMFCC methods have demonstrated:

• Enhanced dynamic range and representational power for high-order cepstral coefficients, eliminating their tendency to cluster near zero (Xu et al., 2018).
• Substantial improvements in downstream classification. In voiceprint recognition for Parkinson’s disease diagnosis:
  • DNN classifiers using WMFCC achieved accuracy rates up to 89.5% for the vowel /u/, outperforming SVMs and conventional MFCCs (Xu et al., 2018).
  • On previously unseen PD data, DNN+WMFCC achieved 100% accuracy for single-vowel classification and 89.1% for multiple vowels.
• In speaker verification benchmarks:
  • Learnable WMFCC frontends reduced equal error rate by up to 6.7% relative on VoxCeleb1 and 9.7% on SITW compared to static MFCC baselines (Liu et al., 2021).
• Improvements extend to accuracy, sensitivity, specificity, Matthews correlation coefficient, and prediction error metrics, suggesting broad enhancement of feature-learnability.

6. Implementation Practices and Regularization

Specific implementation recommendations include:

• Compute entropy-based WMFCC weights per sample with vectorized operations; apply after DCT but before further normalization (Xu et al., 2018).
• Keep frame-level structures intact during weighting to avoid information loss.
• Use BLAS-level matrix routines for efficient calculation throughout FFT, filterbank, DCT, and weighting steps.
• For learnable WMFCC, initialize each kernel from its analytical counterpart and add regularization terms (e.g., toward cosine-shaped windows or orthonormal DCT) to prevent excessive deviation from interpretable forms (Liu et al., 2021).
• Non-negativity (for Mel filterbanks), symmetry (DFT), and orthonormality (DCT) are maintained through explicit constraints or post-update projections.
• In deep learning training, stochastic optimizers such as mini-batch gradient descent are employed, with small batch sizes observed to yield stable convergence in voiceprint tasks.

7. Practical and Research Impact

WMFCC frameworks, both entropy-weighted and learnable, reconcile the rigid structure of classic MFCC-based signal processing with the adaptivity required by contemporary DNN-based classification systems. They enable all cepstral dimensions to participate meaningfully in classification, address representational weaknesses of high-order coefficients, and are broadly applicable in both pathological voice analysis and general speaker verification (Xu et al., 2018, Liu et al., 2021).

These advances underscore the importance of trainable or data-driven frontend preprocessing in speech-related machine learning pipelines, with demonstrated gains even before architectural modifications or augmentation techniques are introduced. The general principle of making MFCCs either explicitly or implicitly weighted is increasingly influential in robust audio representation learning.