metadata
license: mit
language: en
tags:
- pathological-speech
- speech-synthesis
- tts
- voice-conversion
- dysarthric
- ua-speech
Uaspeech Female Dataset
Overview
This dataset contains dysarthric speech samples from a female speaker (F02) in the UA-Speech corpus, prepared for pathological speech synthesis research.
Speaker Information:
- Speaker ID: F02
- Corpus: UA-Speech
- Gender: Female
- Speech Status: Dysarthric
Dataset Statistics
- Total Samples: 1,200
- Total Duration: 1.59 hours
- Sampling Rate: 24,000 Hz
- Format: Audio arrays with transcriptions
Training Split
- Samples: 1,000
- Duration: 1.33 hours
- Avg Duration: 4.8s
- Duration Range: 2.4s - 19.0s
- Avg Text Length: 6 characters
Test Split
- Samples: 200
- Duration: 0.26 hours
- Avg Duration: 4.8s
- Duration Range: 1.9s - 17.8s
- Avg Text Length: 5 characters
Loading the Dataset
from datasets import load_dataset
# Load the dataset
dataset = load_dataset("your-username/uaspeech_female")
# Access train and test splits
train_data = dataset['train']
test_data = dataset['test']
# Each sample contains:
# - 'audio': {'array': numpy_array, 'sampling_rate': 24000}
# - 'text': str (normalized transcription)
# Example usage
sample = train_data[0]
audio_array = sample['audio']['array']
transcription = sample['text']
sampling_rate = sample['audio']['sampling_rate']
Direct Training with Transformers
from transformers import Trainer
from datasets import load_dataset
# Load and use directly with Trainer (no preprocessing needed)
dataset = load_dataset("your-username/uaspeech_female")
trainer = Trainer(
train_dataset=dataset['train'],
eval_dataset=dataset['test'],
# ... other trainer arguments
)