On the Training and Testing Data Preparation for End-to-End Text-to-Speech Application

In the past five years, end-to-end (E2E) text-to-speech (TTS) application has become of interest for many researches related to speech generation from text. This has been the result of the fast-pace development of artificial intelligence and machine learning, not only in image processing, but also in audio and signal processing. Often, the training and testing datasets of TTS applications comprise of thousands of data in forms of text lines and the corresponding audio files which amount up to tens of recording hours. Thus, it is extremely time-consuming for preparing training and testing dataset using manual approach. This work presents an approach for automatic data preparation that is used in Tacotron, Tacotron-2-based Mozilla TTS engine. The well-labeled dataset namely FPT Open Vietnamese Speech Dataset having over 25,000 text lines and recorded audio files is demonstrated in this work. On average, the algorithm takes approximately 45-70 μs to process one text line.