Speech Recognition using TESPAR and Neural Networks

Student: Whee Kian Lee

Supervisor: John Homer

Category: Engineering Thesis Project - Electrical

Speech Recognition has found widespread use in electronic applications. Its appeal lies in its simplicity and together with the ease of operating a device using speech has enormous advantages. However the most modern speech recognition techniques used today have recognition capabilities well below those of a child. These modern techniques are also highly unstable, subjective to noise interferences, require large memory space and complex computations. These factors lead to an increase in production cost of the electronic devices. TESPAR (Time Encoded Signal Processing and Recognition) and Neural Networks technique is the propose approach to speech recognition system in this thesis. TESPAR is a relatively new approach in the field of speech recognition and its coding is based upon approximations to the locations of the real and complex zeros, derived from an analysis of the bandlimited waveforms. The output stream of symbols from the TESPAR coder can be readily converted into a variety of progressively informative fixed dimension TESPAR matrices. TESPAR matrices are ideally matched to the processing requirements of Artificial Neural Networks for which the used of fixed sized training and interrogation vectors is typically essential [1]. The Multi-Layer Perceptron of the Neural Networks is the proposed approach in the classification stage. In the classification stage, each layer of perceptrons is interrogated separately to produce an individual score. An “Accept” or “Reject” decision is made by examining the total score collected. The combination of TESPAR and Neural Network approach results in faster computations, higher accuracy and its architectures can be embodied cheaply in DSP devices. A successful software prototype is being developed to verify the strength of TESPAR and Neural Networks technique. The positive results collected from the assessment stage further substantiate the theories behind TESPAR and Neural Networks.

Thesis Document (PDF)

Other Related Files