Text-to-Speech Systems for Filipino Using Unit Selection and Deep Learning
Edsel Jedd Renovalles, Crisron Rudolf Lucas, Franz de Leon, Angelina Aquino, Izza Jalandoni · 2021
There are several text-to-speech (TTS) systems developed and published for Philippine languages, particularly in Filipino and Cebuano. However, there is still a need to improve the performance of existing systems. Due to the limited amount of linguistic resources and lack of speech data available for Philippine languages, developing a reliable TTS system to support these languages becomes difficult. In this paper, we implement and evaluate the performance of two TTS systems for Filipino. We implemented two methods: unit selection using MaryTTS and deep learning approach using Tacotron-2. We tried applying modification on the F0 contour and duration of the unit selection system and used voice conversion to augment the training data of Tacotron-2. The unit selection system achieved a mean opinion score (MOS) of 3.05, however, the boundary-based F0 modification yielded perceivable distortions in the output and requires more enhancements to become more effective. On the other hand, the use of voice conversion to transform the original multi-speaker data into single-speaker data and producing more samples for training boosted the Tacotron-2 performance from an overall MOS of 1.51 to 2.01.