Evaluating Cognitive Load of Text-To-Speech (TTS) synthesis
| Main Authors: | Govender, Avashna, Valentini-Botinhao, Cassia, King, Simon |
|---|---|
| Format: | Proceeding eJournal |
| Bahasa: | eng |
| Terbitan: |
Deutsche Gesellschaft für Akustik (DEGA e. V.) & RWTH Publications
, 2019
|
| Subjects: | |
| Online Access: |
https://zenodo.org/record/3527987 |
Daftar Isi:
- Current evaluation methods for text-to-speech (TTS) synthesis rely solely on subjective rating scores. Thesetests typically account mostly for how natural or intelligible the voice is. With state-of-the-art systems, thesemeasures are approaching ceiling and therefore alternative measures such as the cognitive load may becomemore meaningful. To our knowledge, there is little or no recent work evaluating the cognitive load of state-of- the-art text-to-speech systems. We use pupillometry as a measure of cognitive load. The pupil has beenfound to dilate upon increased cognitive effort when carrying out a listening task. Currently we are evaluatingspeech generated by a Deep Neural Network TTS synthesiser. In our method, we generate stimuli that stepincrementally from natural speech to synthesized speech by changing only a single feature at a time. Stimuli arepresented to listeners in speech-shaped noise conditions.