Multi Speaker Speech Synthesis System for Indonesian Language
Marvin Jerremy Budiman, Dessi Puji Lestari · 2020
Generally, text-to-speech models only produce voice from a single speaker. The most straightforward method to produce another speaker's voice, is to build a standalone synthesis model for each desired speaker's voice. But such approach needs large amount of training data and computational resource. To overcome the problem, several architectures has been successful in producing synthesized speech from various speakers efficiently in terms of data and computation. One of the architectures is Deep Voice 3. In this work, a multi speaker speech synthesis system is built for Indonesian language. The system is using Deep Voice 3 architecture, with several additional components for preprocessing dan post-processing. Some of the components are specifically implemented for Indonesian language. The system is built using a multi speaker dataset, consists of speech data from 145 Indonesian speaker. This system is evaluated subjectively to assess naturalness, similarity to original speaker, and intelligibility of the produced speech. The result shows that the system has MOS (mean opinion score) of 3.39 for speech naturalness dan 3.11 for speech similarity. In assessing speech intelligibility using SUS (semantically unpredictable sentence), the test gives 73.88% for sentence accuracy and 93.48% for word accuracy.