Phonsiri/thai-elderly-speech

Thai Elderly Speech Dataset (Combined Evaluation Set)

This dataset contains evaluation recordings for Thai elderly speech, combined from Healthcare and Smarthome domains.

Dataset Structure

After extracting Combined_Dataset.zip, the directory structure will look like this:

Combined_Dataset/
├── Train/                        # 80% of the dataset (15,360 files)
│   ├── Accuracy_100/             # Files with 100% baseline accuracy
│   ├── Accuracy_50_99/           # Files with 50-99% baseline accuracy
│   └── All_Dataset/              # Contains ALL Train files combined
├── Test/                         # 20% of the dataset (3,840 files)
│   ├── (all test audio files are stored directly here)
├── train_typhoon_evaluation_results.csv   # Metadata & transcripts for Train set
├── test_typhoon_evaluation_results.csv    # Metadata & transcripts for Test set
└── typhoon_evaluation_results.csv         # Original combined metadata

• Train Set (80%): Used for fine-tuning models. It is subdivided into Accuracy_100, Accuracy_50_99, and All_Dataset to allow flexible targeted training (e.g., training only on highly accurate data vs full dataset).
• Test Set (20%): A completely isolated central test set used for unbiased evaluation across all models. All audio files are located directly in the Test/ folder without subdirectories.
• Metadata Files: The CSV files contain the filenames, reference transcripts, hypothesis transcripts (from the baseline model), and accuracy percentages.