Sistem Pengenal Wicara Menggunakan Mel-Frequency Cepstral Coefficient

Abstract

Human-machine interaction evolves toward a more adaptive and interactive system. There are several media that can be used in human-machine interaction systems, such as voice signals. The process includes converting analog signals into the appropriate meaning, which depend on the noise and reliability of signal characteristic extraction methods. In fact, variations of pronunciation by different people will result in a diversity of voice signal patterns. This research develops technology that can recognize and translate speech according to data that has been trained and can be modified based on user requirement. The voice signal will be separated from the silent signal using voice activity detection. Then, the voice signal is converted to the frequency domain before it is extracted using mel-frequency cepstral coefficients. Cepstral value from MFCC extraction will be identified as words using artificial neural network. This study utilizes a computer with a microphone as a sound recording device and pascal programming language as the basis for building applications. Based on the experimental results, the accuracy is 87% on the speech recognition process with 28 vocabulary sets. Accuracy decreases with more sets of vocabulary. However, the more pronounced speech variations, the greater the accuracy with an average number around 93%.