This rese arch dev oted to t he development of Speec h Re cognition Syst em in Be ngali language t hat works with speaker independent, isolated and subword-unit-based approaches. In our work, the original Bangl a speech words were reco rded an d stored as RIFF (.wa v) file. Th en these wo rds were cla ssified i nto t hree different group s according t o t he number of syllables of t he speec h w ords a nd t hese gr ouping s peech si gnals w ere converted t o digital form, in order to extract features. The features were extracted by the method of Mel Freque ncy Cepstrum Coefficient (MFCC ) a nalysis. Th e reco gnition system i ncludes d irect Euclid ean d istance mea surement techn ique. The test database contained 600 distin ct Bangla speech words and each word was reco rded from six different speakers. The developme nt software is written in Turbo C and c ommon feat ure of t oday's software hav e bee n included. The development system achieved recognition rate at about 96% for singl e spe aker and 84.28% for multiple speakers.