Less errors with TTS? A dictation experiment with foreign language learners

This article reports a contrastive experiment about the use of Text-To-Speech (TTS) synthesis instead of prerecorded utterances in a dictation exercise submitted to students of European Portuguese as a second language (PSL). Forty sentences were extracted from a PSL student book. Twenty of them were synthesized and the other twenty were directly taken from the pre-recorded audio documents of the book. It appeared that the synthetic utterances were easier to transcribe than the prerecorded ones, with word error rates of 26.6% and 33.9% respectively. This result was somehow surprising since the synthetic voice was not built or tuned for learning purposes. Nevertheless, the lower speech rate of the TTS voice (15% slower) may explain this outcome. A manual error categorization showed that less word substitutions and less errors on function words were made on the TTS