AUTOMATIC CHORD ESTIMATION USING COMBINED METHOD OF DEEP LEARNING AND HIDDEN MARKOV MODEL

• Main Author: Budi Ghifari, Januar
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85072
• Institution: Institut Teknologi Bandung

Automatic Chord Estimation (ACE) is a task within the field of Music Information Retrieval (MIR) that aims to automatically identify the sequence of chords in music audio recordings. This thesis proposes the use of a combined method of Deep Learning and Hidden Markov Model (HMM) to improve the accuracy of chord estimation. The research compares feature extraction from audio signals using methods such as Short-Time Fourier Transform (STFT), Constant-Q Transform (CQT), and a combination of both. The extracted features are then used as input for a deep learning model. CNN, as the deep learning model, is used to estimate chords for each window, while HMM is employed as a post-processing method to refine the estimation results from the CNN. This research also focuses on separating chord classification tasks based on chord components to reduce the model's burden in estimating the large number of chord classes. The findings indicate that combining feature extraction methods in audio signal processing and separating classification tasks can improve the accuracy of the ACE system. The segment-based chord symbol recall metric is used to measure chord prediction accuracy. In the single model, the CQT + STFT feature extraction method achieved a chord score of 64.57 without HMM and increased to 65.01 with HMM. In the multi-model approach, the STFT method showed a chord accuracy of 75.54 without HMM.