This paper investigates use of a machine learnt model for recognition of individually words spoken in Urdu language. Speech samples from many different speakers were utilized for modeling. Original time-domain samples are normalized and pre-processed by applying discrete Fourier transformation for speech feature extraction. In frequency domain, high degree of correlation was found for the same words spoken by different speakers. This helped produce models with high recognition accuracy. Details of model realization in MATLAB are included in this paper. Current work is being extended using linear predictive coding for efficient hardware implementation.