Enhancing Amharic Sentence Segmentation with Prosodic Features and Neural Network Models

Rahel Mekonen Tamiru, Hana Mekonen Tamiru, ABEL MULAT ALEMU

Keywords: speech Segmentation, prosody

TL;DR: This study aims to create a sentence-level automatic speech segmentation system for Amharic.

Abstract:

This study focuses on developing a sentence-level automatic speech segmentation system for Amharic. Two approaches were explored. The first approach utilized an automatic tool for segmenting and labeling Amharic speech data, creating an acoustic model through HMM modeling. The system's segmentation was refined using forced alignment AdaBoost techniques. In the second approach, prosodic features
were extracted directly from the speech waveform, and statistical methods including AdaBoost were employed. Additionally, LSTM and Bi-LSTM models were utilized, achieving impressive accuracies of 94.62% and 95.23%, respectively. These approaches contribute to
advancing automatic speech segmentation for Amharic, promising improved accuracy and efficiency.

Submission Number: 12