Abstract
This paper evaluates the performance of 17 machinelearning classifiers in automatically scoring the English proficiency of unconstrained speech. Each classifier was tested with different groups of features drawn from a master set of prosodic measures founded in Brazil’s model [3]. The prosodic measures were calculated from the output of an ASR that recognizes phones instead of words and other software designed to detect the elements of Brazil’s prosody model. The performance of the best classifier was 0.68 (p < 0.01) in terms of the correlation between the computer’s calculated proficiency ratings and those scored by humans. Using only prosodic features, this correlation is in the range of other similar computer programs for automatically scoring the proficiency of unconstrained speech.
Original language | English (US) |
---|---|
Pages (from-to) | 617-620 |
Number of pages | 4 |
Journal | Proceedings of the International Conference on Speech Prosody |
Volume | 2018-June |
DOIs | |
State | Published - 2018 |
Event | 9th International Conference on Speech Prosody, SP 2018 - Poznan, Poland Duration: Jun 13 2018 → Jun 16 2018 |
Keywords
- Automatic speech recognition (ASR)
- Brazil’s prosody model
- Large vocabulary spontaneous speech recognition (LVCSR)
- World Englishes
ASJC Scopus subject areas
- Language and Linguistics
- Linguistics and Language