Identifikasi Voice Menggunakan Jaringan Syaraf Tiruan dan Mel Frequency Cepstrum Coefficients

Abstract

Perkembangan teknologi di bidang pengenalan yang semakin pesat membutuhkan daya fikir, kreatifitas dan inovasi untuk meningkatkan sistem teknologi modern pada bidang ilmu pengetahuan khususnya bidang access. Ilmu yang membahas bidang pengenalan suara dikenal dengan nama pemrosesan sinyal suara digital dan Natural Language Processing. Untuk menyelesaikan permasalahan pengenalan suara voice identification menggunakan pemerosesan sinyal informasi yang menggunakan proses Mel Frequency Cepstrum Coefficients (MFCC) dan menggunakan Jaringan syaraf tiruan (JST) backpropagation dengan proses pembelajaran perubahan bobot. Siklus perubahan bobot atau EPOCH (Exponential Decay) digunakan sebagai inisialisasi pengidentifikasian suara voice pria dan wanita yang diperoleh dengan menggunakan beberapa sampel suara pria dan wanita. Hasil pengenalan voice identification dari gender wanita lebih baik dibandingkan pria dengan nilai error pada pendeteksian suara voice wanita 0.05< Error < 1 lebih kecil dari siklus perubahan bobot jaringan syaraf tiruan backpropagation dengan konstanta laju pelatihan (α=0.2).

The development of information technology requires the power of thinking, creativity and innovation of progressing the increasingly varied life systems to improve modern system technology and sciences of access. The development of voice is known speech and natural language processing. Voice becomes communication medium for human and computer (machine) is used for recognition/identification and access system. One of its application is speech to text application. Some of speech to text research already claimed good accuracy for some languages or voice namely for Robotic, furniture, digital signal processing and biometric technology. For managing the recognition, we used several approaches in this research. The researcher performs information signal processing by using Mel Frequency Cepstrum Coefficients (MFCC) and using artificial neural network (ANN) backpropagation with learning process through weight change. The weight change cycles or EPOCH (Exponential Decay) are used as initialization of voice identification of male and female voice obtained by using multiple sound. The result of voice identification of female gender is better than man with error on detection of female voice 0,05 <error <1 smaller than cycle change of artificial neural network backpropagation with α = 0.2.