Speaker-Independent Digit Recognition Using a Neural Network with Time-Delayed Connections

Abstract

The capability of a small neural network to perform speaker-independent recognition of spoken digits in connected speech has been investigated. The network uses time delays to organize rapidly changing outputs of symbol detectors over the time scale of a word. The network is data driven and unclocked. To achieve useful accuracy in a speakerindependent setting, many new ideas and procedures were developed. These include improving the feature detectors, self-recognition of word ends, reduction in network size, and dividing speakers into natural classes. Quantitative experiments based on Texas Instruments (TI) digit data bases are described.

How Is Time Represented in the Brain?

“…Combined with Hebbian learning schemes, a broad distribution of transmission delays can be used to "concentrate information in time" in that stimuli that were originally spread out over a large time interval are grouped together. This mechanism has also been implemented in several technical applications, such as artificial speech recognition (Unnikrishnan et al 1992).…”

“…For example, the length of single "syllables" of the mating songs of grasshoppers varies up to 300% as a function of the ambient temperature (von Helversen and von Helversen, 1994). Large time warps are, however, also common in human speech where the duration of a word spoken by a given speaker may change by up to 100% depending on the specific circumstance (see, e.g., Unnikrishnan et al 1992). Psychophysical data show that in both systems, even such large variations are tolerated with ease by the receiver.…”