This paper described the process of creating and testing the offline Khmer speech recognition system. This system was created using CMUSphinx with the noise reduction of training audio database, including 85 speakers and 157 words selected from the Khmer language. To evaluate the speech recognition accuracy, there were 100 Khmer transcripts randomly created from the training dictionary for calculating the word and sentence error rate. The recognition accuracy of Khmer speech recognition can achieve up to 89.91% of word recognition accuracy and 90.02% of sentence recognition accuracy.

Keywords : Khmer, phoneme, ASR, CMUSphinx Toolkit, Acoustic Model, Phonetic Dictionary, Language Model.