Isolated-word speech recognition can be divided into two main data reduction problems:

This project focuses on long-term time-variance. The most prevalent method to do this is using HMMs, but it isn’t the only method…

Potentially less computation during operation mode than HMMs (no Viterbi search)

Distributed computation makes them much faster if implemented using specialized hardware

They’re Cool!!

But... they take a long time to train

A neural net with an appropriate design should be able to find these features. But what design?

The network must be time-dependent

One RNN for entire system

One RNN for each word (as when using HMMs)

An Elman network was used (the feedback connection is an exponentially decaying integrator)

The input layer is a gamma-filtered TDNN (time-delay NN) with 10 delay taps and memory depth of 100

80% correct recognition using HW #4 data without added noise (95% correct using HMMs on equivalent inputs)

Obtained on non-optimized and poorly trained networks, due to time constraints

Better results are expected with more training and optimization

The RNNs used are not well suited to classify functions in time directly

Training separate RNNs for each word using prediction works quite well

RNNs could be combined with HMMs or other ANNs for continuous word recognition

Train with noise added to the speech

Apply global optimization to the fixed parameters in the RNN

Try different NN architectures

Try to find a better method of comparing the outputs other than weighted mean squared error (another NN, fuzzy rules, entropy measures, etc.)