Creating Navajo synthetic speech with deep learning

With new developments in machine learning and deep neural networks, speech synthesis technology has made it possible for almost anyone to have a voice, including those with severe speech impairments. However, there exists a mismatch between the technology’s capabilities and the availability of naturalistic and personalized voices in a wide variety of languages, especially indigenous ones. To close this gap, this project seeks to create a natural-sounding speech synthesizer for an adult Navajo speaker with limited training data. A pilot speaker was recorded in a sound-attenuated booth. After removing the false starts, fillers, and other non-speech sounds, the speaker produced just over 2 h of data. The speech synthesis software Tacotron 2 and Waveglow were utilized to produce high-quality synthesized speech. To determine the naturalness of the synthetic speech, listeners participated in a mean opinion score survey and offered qualitative feedback. Our results highlight the feasibility of creating synthetic speech for other low-resource languages. We discuss the viability of exploring more cost-effective means of personalized speech synthesis, as opposed to existing methods for high-resource languages.