English as a Second Language TTS (ESLTTS) dataset

With the progress made in speaker-adaptive TTS approaches, advanced approaches have shown a remarkable capacity to reproduce the speaker’s voice in the commonly used TTS datasets. However, mimicking voices characterized by substantial accents, such as non-native English speakers, is still challenging. Regrettably, the absence of a dedicated TTS dataset for speakers with substantial accents inhibits the research and evaluation of speaker-adaptive TTS models under such conditions. To address this gap, we developed a corpus of non-native speakers' English utterances.

We named this corpus “English as a Second Language TTS dataset ” (ESLTTS). The ESLTTS dataset consists of roughly 37 hours of 42,000 utterances from 134 non-native English speakers. These speakers represent a diversity of linguistic backgrounds spanning 31 native languages. For each speaker, the dataset includes an adaptation set lasting about 5 minutes for speaker adaptation, a test set comprising 10 utterances for speaker-adaptive TTS evaluation, and a development set for further research.