Speech translation (ST) systems consist of three major components: automatic speech recognition (ASR) ,machine translation (MT) and speech synthesis (SS). In most cases the ASR system is tuned by minimizing word error rate (WER). However decreasing WER is not directly guaranteed to improve the translation quality.Because WER only considers the number of word errors, it doesn$B!G(Jt consider the effect of recognition errors on translation. In previous research, ASR and MT have been jointly optimized to improve translation quality [9].Optimization of MT has also used with rich features to improve translation quality [10]. However joint optimization has never been used rich features. In this thesis we jointly optimize the weights using pairwise rank optimization(PRO) [10], which is able to use rich features. We tested the effect of joint optimization using the rich features from MT, ASR, and frequency of recognized words. Experimental result on a travel conversation corpus [28] Showed that the translation quality is not statistical significant difference in PRO and minimum error rate training (MERT) [22]. Rich features do not have an effect of the improving translation quality.