|Cetin, A. Enis - BILKENT UNIV. TURKEY|
|Tewfik, Ahmed - UNIV. OF MN|
Submitted to: Transactions of the ASAE
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: February 1, 2004
Publication Date: March 1, 2004
Citation: Cetin, A., Pearson, T.C., Tewfik, A.H. 2004. Classification of closed- and open-shell pistachio nuts using voice-recognition technology. Transactions of the ASAE. 2004. 47(2):659-664. Interpretive Summary: Voice recognition technologies developed over the past ten years have been able to correctly convert audio signals from speech to text with over 99% accuracy. In addition, these technologies are easy to train for additional applications that they were not initially designed for. However, voice recognition technologies have not yet been applied to agriculture to perform, for example, classifications of good and defective product. This work uses fundamental voice recognition technology to distinguish open and closed shell pistachio nuts based on the acoustic signals created when the nuts strike a hard plate. Results from this method are as good, or better, than results from earlier studies using more conventional signal processing and classification algorithms.
Technical Abstract: An algorithm using speech recognition technology was developed to distinguish pistachio nuts with closed shells from those with open shells. It was observed that upon impact with a steel plate, nuts with closed shells emit different sounds than nuts with open shells. Features extracted from the sound signals consisted of mel-cepstrum coefficients and eigenvalues obtained from the principle component analysis (PCA) of the autocorrelation matrix of the sound signals. Classification of a sound signal was performed by linearly combining the mel-cepstrum and PCA feature vectors. An important property of the algorithm is that it is easily trainable, as are most speech-recognition algorithms. During the training phase, sounds of the nuts with closed shells and open shells were used to obtain a representative vector of each class. During the recognition phase, the feature vector from the sample under question was compared with representative vectors. The classification accuracy of closed-shell nuts was more than 99% on the validation set, which did not include the training set.