Voice Builder: A Tool for Building Text-To-Speech Voices

Voice Builder: A Tool for Building Text-To-Speech Voices Pasindu De Silva2, Theeraphol Wattanavekin1, Tang Hao1, Knot Pipatsrisawat1 1Google,2Google (on contract from Teledirect Pte Ltd) {pasindu,twattanavekin,tanghao,thammaknot}@google.com Abstract We describe an opensource text-to-speech (TTS) voice building tool that focuses on simplicity, flexibility, and collaboration. Our tool allows anyone with basic computer skills to run voice training experiments and listen to the resulting synthesized voice. We hope that this tool will reduce the barrier for creating new voices and accelerate TTS research, by making experimentation faster and interdisciplinary collaboration easier. We believe that our tool can help improve TTS research, especially for low-resourced languages, where more experimentations are often needed to get the most out of the limited language resources. Keywords: text-to-speech, voice building, opensource 1. Introduction There is growing demand for text-to-speech (TTS) voices in various industries as the world moves towards technolo- gies that can interact with users more naturally. Many in- dustries including gaming, film, appliances, and accessibil- ity (Szarkowska, 2011) now have vested interest in creating TTS voices. As a result, there is a lot of needs to produce a variety of good quality voices in different languages. The TTS problem is challenging especially for low-resourced languages, because of the lack of data. Smaller text and au- dio corpora imply that more experiments might need to be done in order to get the most out of the available resources. Tackling this problem for low-resourced languages usually requires close collaboration between computer scientists, who understand the algorithms and know how to utilize the tools, and linguists, who understand linguistic phenomena and can help provide linguistic resources required to im- prove the voices. From our experience building voices for low-resourced lan- guages, we have observed many pain points that prevent these researchers from conducting quick experiments and iterating on their existing baseline voices. These pain points include (i) the complexity of setting up initial voice build- ing infrastructure, which often requires fairly strong tech- nical background to setup and maintain (ii) the heteroge- neous nature of different tools, which requires a non-trivial amount of effort to stitch the required pieces together to form a working pipeline for TTS research. In this paper, we propose Voice Builder, a text-to-speech voice building tool, which comes with a simple web-based user-friendly interface that enables researchers to quickly build and listen to prototype voices and to iteratively im- prove them with ease. 2. Related Work There are many opensource TTS systems available with underlying algorithms ranging from classification regres- sion trees in Festival (Taylor et al., 1998; Black and Taylor, 1997), hidden Markov models in HTS (Zen et al., 2007), to neural networks in Merlin (Wu et al., 2016). Some of these tools depend on components of their predecessors. For example, Merlin requires HTS or Festival to handle the linguistic frontend processing. Some systems, such as MaryTTS (Schr¨oder et al., 2011), try to provide more end- to-end solutions for building various components, such as recording and transcription, required for a complete TTS system. Other tools …

Content truncated to 500 words