Speak & Improve Corpus 2025

Important: This version was prepared by NTNU SMIL to provide the dataset in a Hugging Face Datasets compatible format for convenient use and integration. However, the original dataset was released by third parties, and we do NOT provide direct access.

To use this dataset, you must first visit the official dataset site, agree to the license terms, and download the data from the official source. After obtaining the data, you can use our upload.py script to create your own version on Hugging Face.

Original dataset and license information:

Speak & Improve Corpus 2025 Official Website

Dataset Structure

Data Fields

file_id: Unique identifier for each audio file
audio: Audio recording with preserved original sampling rate
split: Dataset split (train/dev/eval/dev_subset)
speaker_id: Unique speaker identifier
test_part: Test section identifier (P1, P3, P4, P5)
question_number: Question identifier within the test part
question_text: The prompt/question text presented to the speaker
speaking_time: Allocated speaking time in seconds
thinking_time: Allocated thinking time in seconds
transcription: Annotated transcription with linguistic marks
sla_overall_score: Overall spoken language assessment score
sla_P1_score: SLA score for part P1
sla_P3_score: SLA score for part P3
sla_P4_score: SLA score for part P4
sla_P5_score: SLA score for part P5

Transcription Annotations

The transcription field contains annotated transcriptions with linguistic marks when available:

Word-level Marks

# - Disfluency mark (e.g., #dog)
% - Pronunciation mark (e.g., %cat)
* - Partial word mark (e.g., *fish)

For multiple marks, the order is: # then % then * (e.g., #%*word)

Data Splits

Train: Training data for model development
Dev: Development/validation data
Eval: Evaluation data for final testing
Dev Subset: Smaller subset of development data for quick testing

Usage

This dataset is designed for:

Spoken Language Assessment (SLA) research
English proficiency scoring
Speech analysis of second language learners
Audio-based language assessment model training
Transcription analysis with linguistic annotations
Disfluency and pronunciation error detection

Loading the Dataset

from datasets import load_dataset

# Note: You must agree to the original license and create your own version
dataset = load_dataset("your-username/sandi-corpus-2025")

# Access transcription with annotations
for example in dataset["train"]:
    print(f"Audio: {example['file_id']}")
    print(f"Transcription: {example['transcription']}")
    print(f"SLA Score: {example['sla_overall_score']}")

Create Your Own Version

Visit the official website, agree to the license, and download the dataset.
Use our upload.py script to create your Hugging Face version:

python upload.py --repo-name your-username/sandi-corpus-2025

Datasets:

ntnu-smil
/

sandi-corpus-2025

You need to agree to share your contact information to access this dataset