User generated video product reviews in social media is gaining popularity every day due to its creditability and the broad evaluation context provided by it. Extracting sentiment automatically from such videos will help the consumers making decisions and producers improving their products. This paper investigates the feasibility of sentiment detection temporally from those videos by analyzing the transcription generated by a speech recognition system which was not investigated before. Another two main contribution for this paper is introducing a solution to the problem of fixed threshold estimation for the Naive Bayesian classifier output probabilities and irrelative text filtering for improving the sentiment classifcation. Various experiments indicated the proposed system can achieve an F-score of 0.66 which is promising knowing that the sentiment classifier offers an F-score of 0.78 provided that the input text is error free.