NP-TORGO: A Partitioned Dysarthric Speech Dataset

Dataset Overview

NP-TORGO (No Prompt-Overlap TORGO) is a partitioned version of the TORGO dataset designed to address the issue of prompt overlap between speakers. The original TORGO dataset contains dysarthric speech samples from individuals with ALS or CP, but significant prompt overlap can lead to data leakage. NP-TORGO ensures there is no prompt overlap between training, validation, and test sets, providing a more realistic evaluation setting for speech recognition models.

Data Structure

The dataset consists of audio samples and corresponding text transcriptions, with the following key columns:

• session: The recording session of the speaker.
• stringlengths (audio/text): The length of the audio and text samples.
• stringclasses (speaker_id): The unique identifier of the speaker.
• test_data: The split indicator, where:
  • 1 denotes the test set.
  • 0 denotes the training and validation set.
• file_path: The path to the corresponding audio file.
• transcription: The ground truth text for the audio sample.

Data Filtering

NP-TORGO applies the following filtering logic to ensure a clean separation between training, validation, and test sets:

Data Splitting Strategy

• If test_data == 1, the sample belongs to the test set.
• If test_data == 0, the sample is part of the train or validation set.

Speaker Partitioning Strategy

To ensure a fair evaluation, NP-TORGO follows a predefined rule for speaker assignment:

This means:

• If the test speaker is not F03, then the validation speaker is F03.
• If the test speaker is F03, then the validation speaker is set to F04.

This approach prevents overlap and ensures a robust validation protocol.

Citation

If you use NP-TORGO in your research, please cite our paper:

Enhancing AAC Software for Dysarthric Speakers in e-Health Settings: An Evaluation Using TORGO
Macarious Kin Fung Hui, Jinda Zhang, Aanchan Mohan
IEEE ICC '25