Template Based Real-Time Speech Recognition Using Digital Filters on DSP-TMS320F28335

The focus of this paper is to implement speech recognition on a DSP based embedded system and increase its accuracy of correct responses for real-time standalone applications. Template matching algorithm was used to recognize a speaker, based on the stored characteristic features of his/her unique utterance in every particular. Features were derived from the energy component obtained at the output of the digital IIR filters. For this purpose, a filter bank of 15 elliptical filters was designed for 15 different frequency ranges. Each of them yielding a response for the frequency it was designed for and filtering out other frequencies. Furthermore, statistical features were introduced to add contrast between similar sounding words. This approach enabled us to record a high accuracy rate, making it viable to be used in real time low-cost applications.