Evaluating WaveNet Synthetic Speech for English as Second Language Listening Activities

Practitioners in language education believe learners can get into a further stage of language acquisition by understanding a slightly more advanced linguistic input than their current level. This situation leads many researchers to create a learning tool centred around listening activities. However, there is a demand for diverse and massive personalized audio input to build such a great learning tool. To deal with this issue, researchers in the past have shown the potentiality of Text-to-Speech (TTS) technology. Previous research showed that various TTS-based learning activities had achieved promising results. Although TTS implementation has achieved excellent results, only a few studies have formal evaluations on the speech quality of TTS, particularly for English as Second Language (ESL). This research tried to evaluate the speech quality of generated audio materials from a state-of-the-art TTS system called WaveNet. The speech quality was measured by comparing the generated materials’ pronunciation accuracy, comprehensibility, intelligibility, and naturalness with a human voice. Our experiment results showed that synthetic speech had a lower favourable rating for listening comprehension activities compared to human speech. However, as WaveNet produced audios are easier to understand, we believe they can enrich students’ audio input in a less than ideal situation where native speakers are difficult to find.