2014 12th International Conference on Frontiers of Information Technology (FIT) (2014)
Dec. 17, 2014 to Dec. 19, 2014
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/FIT.2014.72
This paper revealed the analysis of speaker independent isolated Pashto spoken numbers for determination of automatic speech recognition. Initially the database was developed, the database encompasses isolated Pashto numbers from sefer (0) to sul (100). Fifty speakers (25 male, 25 females with different ages) that can frequently speak yousafzai dialect were selected for recording. The recording has been done under acoustically controlled environment by the use of Sony PCM-M 10 Linear Recorder and saves in. Wav format. Then, by exploiting Adobe Audition ver. 1.0 the recorded audio files of Pashto numbers were splitted to get isolated Pashto numbers. The recognition experiment was accomplished for first 50 numbers of the database spoken by 50 speakers for the first time up to the author knowledge. Speech signal were analyzed to extract feature vector through MFCC, while LDA were used to probe the difference between the groups to identify various kind of speech classes. The result had shown 83% recognition accuracy on training data and 80% on testing data. The analysis of results has also been discussed in detail.
Mel frequency cepstral coefficient, Speech recognition, Databases, Feature extraction, Accuracy, Speech, Testing
Tanzeela, A. W. Abbas, Z. Ali and B. Uddin, "Analyzing the Impact of MFCC and LDA for the Development of Isolated Pashto Spoken Numbers ASR," 2014 12th International Conference on Frontiers of Information Technology (FIT), Islamabad, Pakistan, 2014, pp. 350-354.