LEVERAGING PAUSE DETECTION FOR ENHANCED TTS DATASET GENERATION

Abstract

This paper presents a novel tool designed to facilitate the creation of datasets for 
Text-to-Speech (TTS) systems. Current methods for generating these datasets often involve 
labor-intensive manual segmentation and transcription of audio data. The proposed tool 
addresses this challenge by utilizing pause detection to automatically segment the audio 
input, such as audiobooks into manageable segments, typically around 15 seconds in length. 
The average pause duration is computed to guide the segmentation process. The tool also 
provides a user-friendly interface where the segmented audio and its corresponding tran
scription are displayed. Users can play the audio segments and mark the corresponding text, 
which is then highlighted in green and saved as an entity pair in a metadata.csv file. Evalua
tion demonstrates that this tool can significantly streamline the process of creating high
quality TTS datasets, reducing the time required by up to 60% compared to manual meth
ods. To assess the effectiveness of the tool, an dataset was created for Armenian language, 
and compared with the Common Voice dataset by training two Tacotron 2 models with 
WaveNet vocoders. Despite a longer training time, due to higher quality audio, the model 
trained on our dataset showed superior performance in terms of speech naturalness (MOS 
3.7 vs. 3.2). This quality improvement has the potential to enhance the performance of TTS 
systems, particularly for low-resource languages.