Analyzing Word Error Rate using String Metrics Algorithm

Abstract

The android speech recognizer acquires speech at run time through a microphone and processes the sampled speech to recognize the uttered text. But the output texts do not match with users’ data because English and Myanmar texts are different. This system therefore proposed a method for obtaining more detail about actual translation errors in the generated output by using the Word Error Rate (WER) based on the string matching algorithms. This paper has investigated string metrics and compared the performance of edit distance like Leveshtein Distance (LD), Q-gram, cosine similarity and dice coefficient by conducting an experiment on the Myanmar name using android speech recognizer output. In order to better fit a variety of android recognition problem over strings, using the edit distance of LD is considered to be an appropriate approach. This paper shows over several experiments that the distance obtains goods results in comparison with other normalized edit distance.