A Hybrid Attention Mechanism to Improve Tacotron 2 Performance for Indonesian Text-to-Speech Synthesis

This paper presents a speech synthesis system for the Indonesian language using the Tacotron 2 architecture and HiFi-GAN vocoder. While Tacotron 2 has demonstrated strong performance in high-resource languages, adapting it for Indonesian remains underexplored. To address this, we propose a hybrid attention mechanism that combines location-sensitive and content-based attention to improve alignment accuracy and convergence during training. The system is trained on a custom Indonesian audiobook dataset prepared in LJSpeech format. We compare several model variants, including versions with and without phoneme conversion, and evaluate them using both subjective and objective metrics. Speech quality is assessed through Mean Opinion Score (MOS) ratings, while model performance is evaluated via attention weight visualization. Results show that the proposed hybrid attention mechanism accelerates alignment learning and produces speech with improved naturalness and intelligibility, confirming its effectiveness for low-resource TTS in Indonesian.