PERBANDINGAN AKURASI EKSTRAKSI FITUR MEL-SPEKTROGRAM DAN MFCC DALAM PENGENALAN BUNYI KONSONAN BILABIAL MENGGUNAKAN CNN

Putra Andika Meifani, NIM.: 21106050031 (2026) PERBANDINGAN AKURASI EKSTRAKSI FITUR MEL-SPEKTROGRAM DAN MFCC DALAM PENGENALAN BUNYI KONSONAN BILABIAL MENGGUNAKAN CNN. Skripsi thesis, UIN SUNAN KALIJAGA YOGYAKARTA.

	Text (PERBANDINGAN AKURASI EKSTRAKSI FITUR MEL-SPEKTROGRAM DAN MFCC DALAM PENGENALAN BUNYI KONSONAN BILABIAL MENGGUNAKAN CNN) 21106050031_BAB-I_IV-atau-V_DAFTAR-PUSTAKA.pdf - Published Version Download (6MB)
	Text (PERBANDINGAN AKURASI EKSTRAKSI FITUR MEL-SPEKTROGRAM DAN MFCC DALAM PENGENALAN BUNYI KONSONAN BILABIAL MENGGUNAKAN CNN) 21106050031_BAB-II_sampai_SEBELUM-BAB-TERAKHIR.pdf - Published Version Restricted to Registered users only Download (11MB) | Request a copy

Abstract

The recognition of bilabial consonants is one of the areas of speech signal processing that can be applied to Automatic Speech Recognition (ASR) systems. Speech feature representations play an important role in determining the ability of a model to perform classification. This study aims to analyze and compare the performance of Mel-Spectrogram and Mel-Frequency Cepstral Coefficients (MFCC) feature representations in recognizing the bilabial consonants /p/, /b/, and /m/ using a Convolutional Neural Network (CNN). The research method consists of speech recording data collection, audio preprocessing, feature extraction using Mel-Spectrogram and MFCC, dataset partitioning using Stratified Sample-Level Partitioning and Speaker-Level Partitioning, CNN model training, and model evaluation. Each model was evaluated using 24 test data with accuracy and Root Mean Square Error (RMSE) as evaluation metrics. Both feature representations were tested under each dataset partitioning scenario to determine their performance differences in bilabial consonant classification. The results show that Mel-Spectrogram achieved higher accuracy than MFCC under both dataset partitioning scenarios. The best accuracy was achieved by Mel-Spectrogram under Speaker-Level Partitioning, with an accuracy of 83.33%, while the best RMSE was achieved by MFCC under Speaker-Level Partitioning, with an RMSE of 0.291898. Under Stratified Sample-Level Partitioning, Mel-Spectrogram achieved an accuracy of 75.00% with an RMSE of 0.362090, while MFCC achieved an accuracy of 70.83% with an RMSE of 0.344311. Therefore, Mel-Spectrogram demonstrated better performance in terms of accuracy, whereas MFCC produced a lower RMSE. In addition, Speaker-Level Partitioning resulted in better evaluation performance than Stratified Sample-Level Partitioning for both feature representations in the conducted experiments.

Share this knowledge with your friends :

Actions (login required)

View Item