Development of Speech recognition technique for Marathi numerals using MFCC & LFZI algorithm Deepali Malewadi Prof. Gouri Ghule Vishwakarma Institute of Information Technology Vishwakarma Institute of Information Technology Pune, India Pune, India deepalimalewadi@gmail.com gouri.ghule@viit.ac.in Abstract ─ India is a multilingual country, so a very less amount of work is done on the Regional languages. Marathi is one of the important regional language of Maharashtra. New interface system with high precision are in progress of development by researchers. This paper presents Speech recognition by collection of database of isolated Marathi numerals ranging from zero ( shunya ) to nine ( nau) which is implemented using two feature extraction techniques Mel Frequency Cepstral Coefficient (MFCC) and Low pass Filter Zero Interpolation (LFZI). The uttered speech samples of Marathi numerals are recorded of both male and female for about 1 sec of duration. A database of total 1000 samples is collected and preprocessing is done on each sample and further implementation that is feature extraction and classification is carried out by the MATLAB. LFZI is one of the discrete wavelet transform method. It has numerous applications such as bank cheque processing, passport number, postal zip code, for physically impaired people. Keywords: Speech Recognition; Marathi numerals; Mel filter Cepstral coefficients; Discrete wavelet transform; Low pass filter Zero interpolation. I. INTRODUCTION Speech is the most desirable medium of communication among humans. During speech processing, entire processing of speech is not possible therefore the recorded speech sample is initially fragmented into small frames. Thus processing can now take place frame wise. A raw speech cannot be directly manipulated, firstly it is preprocessed were the voiced part and the unvoiced or silence part are separated totally, in order to avoid mismatch of samples during recognition. Speech recognition basically means talking to a computer, then recognizing what is been said and finally making into a real time system. An automatic Speech recognition (ASR) system uses two well-known phases: feature extraction and classification, where feature extraction is used at the front-end for speech in order to convert the recorded waveform to some form of acoustic representation known as feature vectors. The features are based on time-frequency representation of acoustic signals, which are computed at regular intervals (e.g., every 10ms). And State vector machine (SVM) covers the back-end for pattern classification. A numerous work is done in non-Indian languages, Marathi is one of the important regional language of India considering speech not as great work is established in Marathi Language. The underlying objective of this paper is to recognize Marathi numerals ranging from zero to nine with the help of two feature extraction techniques Mel frequency Cepstral coefficient (MFCC) and Low pass filter and zero Interpolation (LFZI). LFZI is one of the discrete wavelet transform method which provides better performance. Thus the database contains samples of both male and female for about 1 second with a sampling frequency of 8000 Hz and also the speech samples are recorded with the help of MATLAB. The paper is organized as follows: Literature survey is included in Section II, Methodology adopted is presented in Section III, Experiments carried out till date and their results are included in Section IV, Future work and results found till date are concluded in Section V.