yuwchen/OpenEnded

This is the official dataset introduced in the paper.

OpenEnded: An Open-Response Speech Corpus for Speaking Proficiency Assessment with Human Annotations and ALM Supervision

Speech recordings of English-as-a-foreign-language learners with utterance-level speaking-assessment (ASA) scores: Accuracy, Fluency, Prosody.

Contents

wav/             9,782 WAV files (~8.6G)
csv/
  gemini2_train.csv       
  gemini2_dev.csv
  human_test.csv    
  meta_data.csv

Splits

Split	File	# of utterances	# of speakers	.
train	gemini2_train.csv	6,109	698	The labels are pseudo-labels generated by the audio-language model (i.e., Gemini 2.0)
development	gemini2_dev.csv	2,673	571	The labels are pseudo-labels generated by the audio-language model (i.e., Gemini 2.0)
test	human_test.csv	1,000	250	The labels are human-annotated and include the individual scores from all three annotators, as well as their average score.
total		9,782	1,003 distinct

Note:

• test is speaker-disjoint from both train and validation (0 overlap)
• train and development share 516 speakers. Development is intended as an in-domain held-out signal for early stopping, not for measuring speaker generalization.
• Unique speakers are identified based on anonymous speaker IDs provided by the language learning system. Because the data were collected anonymously, we cannot rule out the possibility that the same speaker may have multiple IDs.

Score scale

accuracy, fluency, prosody are integers in [1, 5], higher = better.

Citation

@inproceedings{chen2026openended,
title = {OpenEnded: An Open-Response Speech Corpus for Speaking Proficiency Assessment with Human Annotations and ALM Supervision},
author = {Yu-Wen Chen, Eric Zhou, Evelyn Ding, Tianyi Shen, Zhou Yu, Julia Hirschberg},
booktitle = {Proc. SLT 2026}
}