Comparison Between Two Spatio-Temporal Organization Maps for Speech Recognition

Zouhour Neji Ben Salem, Laurent Bougrain, Frédéric AlexandreLecture Notes in Computer Science 2006 · DOI · OpenAlex

Abstract

This paper evaluates two biologically inspired extensions of the Self-Organizing Map (SOM) designed to handle temporal data: the ST-Kohonen and the Time-Organized Map (TOM). By testing these models on speech recognition tasks, the authors demonstrate that while both bio-inspired approaches perform well, the TOM model achieves superior accuracy by mimicking the way waves of activity propagate through the brain's cortex.

Introduction

Automatic Speech Recognition (ASR) is a classic spatio-temporal problem. Speech signals are sequences where timing and order are as critical as the spectral content. While traditional tools like Hidden Markov Models (HMMs) dominate the field, biologically inspired neural networks offer an underutilized alternative. Standard Artificial Neural Networks (ANNs) often struggle with the highly variable time dimension of speech. This study explores models that explicitly encode temporal information using neurophysiological principles, such as neuron firing delays and intervals.

Methodology

The researchers compared two distinct methods for integrating temporal information into a Self-Organizing Map:

ST-Kohonen: Complex Number Encoding

The ST-Kohonen model uses the domain of complex numbers to encode data. Each input is assigned a complex number where the magnitude represents signal amplitude and the phase represents temporal delay. This provides the network with two degrees of freedom to process space and time simultaneously. A "Hermitienne distance" is employed to identify the winning neuron during training.

Figure 1: Encoding of inputs using complex numbers in the ST-Kohonen model, where phase represents temporal delay.

Time-Organized Map (TOM): Wave Propagation

Instead of complex numbers, TOM utilizes the Inter-Spike Interval (ISI)—the time gap between consecutive stimuli. The core innovation is the use of "propagating waves." When a stimulus reaches the map, it generates a wave of activity that travels across the neural layer. The interaction between these waves translates temporal distances into physical spatial distances on the map. A key contribution of this work is extending the TOM algorithm from a one-dimensional model to a two-dimensional map.

Figure 2: The mechanism of wave propagation and shifting in the TOM neural layer, illustrating how temporal intervals are mapped to spatial coordinates.

Experimental Results

The models were evaluated using the TI-Digit database, a standard benchmark for isolated word recognition, featuring 110 speakers and eleven vocal digits. Raw audio was converted into 12-coefficient Mel Frequency Cepstral Coefficients (MFCC).

Testing was conducted across three protocols:

  • Monolocutor: Same speaker for training and testing.
  • Speaker Independent: Training on a set of speakers, testing on others.
  • Unknown Locutor: Testing on speakers not seen during training.
Protocol ST-Kohonen Accuracy TOM Accuracy
Monolocutor 98.56% 99.34%
Speaker Independent 97.5% 98.32%
Unknown Locutor 96.9% 98.45%

TOM consistently outperformed ST-Kohonen, likely due to its ability to capture finer spatio-temporal structures through wave-interaction mechanisms.

Discussion and Limitations

The primary drawback of the TOM model is its computational efficiency. The algorithm is slow to converge, requiring approximately $10^9$ iterations. This inefficiency stems from the algorithm updating only the "winner" neuron in each loop, unlike the more efficient ST-Kohonen approach.

Future work will focus on testing these models in noisy real-world environments and integrating them with other biological simulations, such as human cochlea models, to create a fully bio-inspired speech recognition pipeline.

Summary updated

BibTeX
@inproceedings{salem2006comparison,
  author = {Zouhour Neji Ben Salem and Laurent Bougrain and Frédéric Alexandre},
  title = {{Comparison Between Two Spatio-Temporal Organization Maps for Speech Recognition}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2006},
  doi = {10.1007/11829898\_2},
}
Figure