Good approach is to use 120 sample frame with 80 sample overlap and SF=8khz Reffer ETSI Distributed Speech Recognition Frontend to understand preprocessing required. Its here http://webapp.etsi.org/workprogram/Frame_WorkItemList.asp?qETSI_STANDARD_TYPE=%27ES%27&qETSI_NUMBER 1108&qTB_ID=&butExpertSearch=++Search++ Besta luck ~rAGU _______________________________ From: "Gina Upperman" Subject: Simple Speech Recognition Program with LPC Hi All - I am trying to make a simple speech recognition program in Code Composer Studio for a class project, and it is so far not much better than random. I was wondering if anyone had any suggestions on my current approach or knew of a good reference on a basic signal processing algorithm. I am a beginner to speech processing, so any advice would be greatly appreciated. My current approach is to input a floating-point representation of a wav file containing a single recorded word sampled at 4 kHz. I window it into 75 ms frames, and I remove any of these frames where the majority of the speech values are a very small number (ie, to eliminate the frames that occur before and after you start and stop speaking the word). I then calculate and store the LPC coefficients for each of these frames. I do this for approximately 20 different words, which will form the codebook of possible words. Then I input the word to be recognized from a different speaker, using the same method, and compare it to each of the words in the codebook to find the best possible match. For each possible word, I take the sum of the squares of the differences in the LPC coefficients for the codebook word and word to be identified, in each of the 75 ms frames. I take the word with the least total difference to be the match. This approach is not accurately recognizing the speech inputs, and I would very much appreciate any suggestions. Thank you, Gina Upperman

Hi All - I am trying to make a simple speech recognition program in Code Composer Studio for a class project, and it is so far not much better than random. I was wondering if anyone had any suggestions on my current approach or knew of a good reference on a basic signal processing algorithm. I am a beginner to speech processing, so any advice would be greatly appreciated. My current approach is to input a floating-point representation of a wav file containing a single recorded word sampled at 4 kHz. I window it into 75 ms frames, and I remove any of these frames where the majority of the speech values are a very small number (ie, to eliminate the frames that occur before and after you start and stop speaking the word). I then calculate and store the LPC coefficients for each of these frames. I do this for approximately 20 different words, which will form the codebook of possible words. Then I input the word to be recognized from a different speaker, using the same method, and compare it to each of the words in the codebook to find the best possible match. For each possible word, I take the sum of the squares of the differences in the LPC coefficients for the codebook word and word to be identified, in each of the 75 ms frames. I take the word with the least total difference to be the match. This approach is not accurately recognizing the speech inputs, and I would very much appreciate any suggestions. Thank you, Gina Upperman