Exploratory Feature Extraction in Speech Signals

Nathan IntratorNeurIPS 1990

Introduction

In high-dimensional data such as speech signals, researchers consistently face the "curse of dimensionality." To make classification computationally feasible and robust, it is necessary to reduce the number of features. Standard unsupervised methods like Principal Component Analysis (PCA) search for variance, but high variance does not always equate to "interesting" or discriminative information.

Conversely, supervised methods like back-propagation reduce dimensionality by focusing on features that minimize error on a specific training set. However, these models often become "lazy." They tend to latch onto the most obvious statistical shortcuts—such as a specific high-frequency burst—that may not generalize to different contexts, like a new speaker or a shift from an unvoiced consonant (e.g., "p") to a voiced one (e.g., "b"). This paper introduces an unsupervised neural network based on biological learning rules (BCM theory) that performs dimensionality reduction by seeking structurally and linguistically meaningful projections of high-dimensional data.

Methods: BCM Theory and Projection Pursuit

The author proposes an unsupervised approach rooted in Exploratory Projection Pursuit and BCM (Bienenstock, Cooper, and Munro) theory.

  1. Seeking Non-Gaussianity: According to statistical theory, most low-dimensional projections of high-dimensional data clouds resemble a normal (Gaussian) distribution. Therefore, the most "interesting" features are typically found in projections that deviate significantly from a Gaussian distribution—specifically, those that are multimodal (containing multiple distinct clusters).
  2. The BCM Neuron: The model utilizes synaptic modification equations originally designed to explain visual cortical plasticity. These neurons employ a dynamic, non-linear threshold ($\Theta_m$) that adjusts based on average network activity.
$\phi$ and Loss Functions in BCM Theory

This dynamic threshold mechanism prevents the data distribution from collapsing into a single point. Instead, it "pushes" the synaptic weights to discover directions where the data naturally splits into distinct modes.

  1. Lateral Inhibition: To extract multiple features simultaneously, the architecture uses a network of mutually inhibitory neurons. This lateral inhibition forces different neurons to "specialize" in distinct features of the data rather than converging on identical representations.
  2. Non-linear Refinement: To handle outliers and facilitate "symmetry breaking," the model incorporates sigmoidal non-linearities, making the feature extraction process far more robust than purely linear projection methods.

Experimental Setup

The study evaluates this approach on the recognition of six stop consonants: the unvoiced [p, k, t] and the voiced [b, g, d].

  • Input Space: The data consists of 440 dimensions, derived from 20 time windows of 32ms each, with 22 energy levels per window (based on Zwicker critical band filters).
  • Speakers: Audio data was collected from three native American English speakers (two male, one female).
  • Training/Testing Split: To rigorously test generalization, the dimensionality reduction was trained only on the unvoiced tokens of a single male speaker. The resulting extracted features were then tested on both voiced and unvoiced tokens across all three speakers.

Key Results and Insights

The BCM-based feature extraction consistently and significantly outperformed standard back-propagation (B-P) when generalizing to new, unseen conditions:

Task / Speaker Back-Prop (B-P) Accuracy BCM Accuracy
BSS (Unvoiced /p,k,t/) 100% 100%
BSS (Voiced /b,g,d/) 83.4% 94.7%
LTN (Voiced /b,g,d/) 78.3% 93.2%
JES (Both - Female) 88.0% 99.4%

The "Lazy" Back-prop Problem vs. Structural Discovery

In this study, back-propagation focused almost entirely on high-frequency bursts at the start of consonants. While this yielded perfect accuracy on the training set, it failed significantly when voicing changed the low-frequency profile of the sound.

Because the BCM network is unsupervised and explicitly looks for multimodality, it does not "know" the labels. It is forced to find the inherent clusters in the data. This leads it to discover "linguistically meaningful" structures—such as formant tracking and burst locations—which are naturally more invariant across different speakers. The BCM network's ability to maintain nearly 100% accuracy on a female speaker (JES) after being trained exclusively on a male speaker's unvoiced stops is a powerful testament to its robustness. Furthermore, it demonstrates that biologically plausible equations modeled after the visual cortex are surprisingly effective at solving abstract statistical problems in audio signals.

Limitations and Future Work

While highly effective, the approach has notable limitations:

  • Outlier Sensitivity: Although non-linearities were introduced to mitigate this, projection pursuit methods are historically sensitive to outliers, which can artificially "fake" multimodality.
  • Computational Complexity: Finding a general multimodal measure is computationally demanding when the underlying structure is unknown, requiring careful tuning of the loss functions.
  • Scope: The current study is limited to stop consonants followed by the vowel [a]. Future research must address more complex phonetic environments and continuous speech.

Ultimately, the paper suggests a hybrid future: while supervised learning excels at classification, the critical feature extraction phase may be best handled by unsupervised, exploratory methods like BCM.

Summary updated

BibTeX
@inproceedings{intrator1990exploratory,
  author = {Nathan Intrator},
  title = {{Exploratory Feature Extraction in Speech Signals}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1990},
}
Figure