Reproduce “Speak & Improve Corpus 2025: an L2 English Speech Corpus for Language Assessment and Feedback” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kate Knill, Diane Nicholls, Mark J.F. Gales, Mengjie Qian, Pawel Stroinski
Speak & Improve Corpus 2025 releases approximately 315 hours of spontaneous L2 English speech from learners at roughly A2–C1 proficiency, paired with holistic scores, disfluent transcripts, fluent versions, and grammatical corrections. The resource supports a shared evaluation pipeline spanning non-native ASR, proficiency assessment, disfluency detection, spoken grammatical error correction, and learner feedback, while its non-speaker-disjoint partitions and uneven speaker-background coverage limit some forms of generalization analysis.
Key insights
- Separating transcription from correction preserves the distinction between what a learner said, how disfluencies can be removed, and how grammatical errors should be repaired.
- Open speaking tasks expose systems to communicative language and spontaneous phenomena that are largely absent from read-speech pronunciation corpora and written correction datasets.
- The corpus provides broad task and learner variation, but its intermediate-level concentration, incomplete L1 balancing, exclusion criteria, and possible speaker overlap across splits must be considered when interpreting results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.