A connectionist approach to rate adaptation

Abstract

In previous work [2], we presented a technique for automatically adapting to the rate of an incoming signal. The idea is as follows: First, a model of the signal is built using a recurrent neural network trained to predict its input at some delay, for a "typical" rate of the signal. Then, with the weights fixed, the time constant of the network is adapted using gradient descent. The time constant controls the processing speed of the network; thus, by adjusting the time constant, we adapt the network's processing rate to match that of the input signal.This method has previously been applied to several signals: sets of sine waves differing in frequency and in phase, a multidimensional signal representing the walking gait of children, and the energy contour of a simple speech utterance. This paper describes experiments on a voicing distinction task using simulated speech data. Results indicate that, using data at only a few rates, our rate adaptation technique can be used to train a system which can then adapt to other rates not represented in the training set. Furthermore, the network's time constant appropriately reflects rate changes in the input signal and can be used as a measure of the signal's rate. We will discuss how this technique can be used to build speech systems that are robust to speaking rate variation.