QUALITY IMPROVEMENT OF LOW-RESOURCED BAHASA INDONESIA EXPRESSIVE SPEECH SYNTHESIS USING CROSS-LINGUAL TRANSFER LEARNING AND TACOTRON2

In this paper, we aim to develop TTS systems for a low-resource language. We proposed an alternative approach that enables effective construction via recurrent sequence-to-sequence feature-prediction networks and knowledge transfer from a low-resource language to improve speech quality. We have investigated fine-tuning the English-pretrained Tacotron2 model on a limited corpus of expressive Bahasa Indonesia speech, based on phonetically balanced data, to synthesize natural speech. We use three expressive styles - happiness, sadness, and anger-each with four male and four female speakers. The performance of the dual method is compared with that of the hidden Markov model-based speech synthesis system (HTS). A perceptual quality in speech test using the Virtual Speech Quality Objective Listener Mean Opinion Score (ViSQOL MOS) method earned the highest value for FYAT voice in HTS, with a score of 3.13/5.00 for happiness, 3.14/5.00 for anger, and 4.05/5.00 for sadness. Then, for the proposed technique, the highest score was 3.94/5.00 for happiness, 4.07/5.00 for anger, and 3.77/5.00 for sadness. It is observed that the proposed technique outperforms the conventional method, likely due to the shorter duration of the speech data used-only 30 minutes.