med-cre / README.md
jsbeaudry's picture
Add 15 files
dde22c5 verified
---
license: mit
task_categories:
- automatic-speech-recognition
- text-to-speech
tags:
- speech
- audio
- haitian_creole
- healthcare
- human
- multilingual
language:
- ha
size_categories:
- n<1K
pretty_name: Multi-Domain Haitian_creole Speech Dataset
---
# Multi-Domain Haitian_creole Speech Dataset
This dataset contains 12 audio recordings with corresponding text transcriptions across multiple languages and domains.
## Dataset Description
A comprehensive collection of audio files paired with text transcriptions, featuring both synthetic and natural speech across various domains. Suitable for automatic speech recognition (ASR), text-to-speech (TTS), and domain-specific speech processing tasks.
## Dataset Structure
Each entry contains:
- `id`: Unique identifier (UUID)
- `text`: Transcription text in the specified language
- `audio`: URL to the audio file (with AWS S3 signed URLs)
- `nature`: Type of audio (e.g., "synthetic", "natural")
- `language`: Language of the audio/text
- `domain`: Domain/topic category (e.g., "agriculture", "healthcare", "education")
## Languages
This dataset includes the following languages:
- **Haitian_creole** (ha): haitian_creole
## Domains
Content spans across multiple domains:
- **Healthcare**: Domain-specific terminology and context
## Audio Nature
The dataset includes different types of audio:
- **Human**: Natural human speech
## Usage
```python
from datasets import load_dataset
import json
import requests
from io import BytesIO
import pandas as pd
# Load using datasets library
dataset = load_dataset("jsbeaudry/med-cre")
# Or load JSON directly
with open("dataset.json", "r", encoding="utf-8") as f:
data = json.load(f)
print(f"Dataset contains {len(data)} audio-text pairs")
# Create DataFrame for analysis
df = pd.DataFrame(data)
print("\nDataset breakdown:")
print(f"Languages: {df['language'].value_counts().to_dict()}")
print(f"Domains: {df['domain'].value_counts().to_dict()}")
print(f"Nature: {df['nature'].value_counts().to_dict()}")
# Filter by criteria
swahili_agriculture = [item for item in data
if item['language'] == 'swahili' and item['domain'] == 'agriculture']
print(f"\nSwahili agriculture samples: {len(swahili_agriculture)}")
# Example: Download and process audio
def download_audio(url):
response = requests.get(url)
return BytesIO(response.content)
# Get first audio file
audio_data = download_audio(data[0]['audio'])
print(f"Audio downloaded for: {data[0]['text'][:50]}...")
```
## Sample Data
```json
{
"id": "2317776c-d722-4870-bcfa-99b3b58526c4",
"source": "Jean",
"text": "Kijan ou santi ou jodi a, paske pran swen tèt ou enpòtan anpil pou sante mantal ou.",
"audio": "https://voiceovers-haiti.s3.us-east-2.amazonaws.com/0e230ff9-f13e-4d3b-be7d-fb3b19977511_d28cfdee-75e0-489a-a790-49db5343dd8a_human.wav?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=AKIAXF2BWYGA2CKJX4LH%2F20251005%2Fus-east-2%2Fs3%2Faws4_request&X-Amz-Date=20251005T063333Z&X-Amz-Expires=3600&X-Amz-Signature=db472454266f6ace8db4563a71a3705ba9e49635f17a97217c4975ef9596e07e&X-Amz-SignedHeaders=host&x-amz-checksum-mode=ENABLED&x-id=GetObject",
"nature": "human",
"language": "haitian_creole",
"domain": "healthcare"
}
```
## Sample Transcriptions by Domain
### Healthcare Domain
1. **Haitian_creole** (human):
"Kijan ou santi ou jodi a, paske pran swen tèt ou enpòtan anpil pou sante mantal ou."
*ID*: `2317776c...`
2. **Haitian_creole** (human):
"Si w santi estrès oswa tristès, pa pè chèche sipò yon pwofesyonèl nan sante mantal, yo la pou ede w."
*ID*: `9665761e...`
## Dataset Statistics
- **Total audio files**: 12
- **Languages**: 1 (Haitian_creole)
- **Domains**: 1 (healthcare)
- **Audio types**: human
- **Average text length**: 74 characters
- **Audio hosting**: voiceovers-haiti.s3.us-east-2.amazonaws.com
### Distribution by Category
| Category | Values |
|----------|---------|
| Haitian_creole | 12 samples |
| Healthcare | 12 samples |
| Human | 12 samples |
## Audio Format
Audio files are stored locally in the dataset as WAV files. When loaded with the datasets library, audio is automatically converted to the standard format:
- **Format**: WAV
- **Sampling Rate**: Preserved from original (typically 16kHz or 22kHz)
- **Channels**: Mono
- **Bit Depth**: 16-bit or 32-bit float
- **Access**: Direct array access via `dataset['train'][index]['audio']['array']`
## Use Cases
This dataset can be used for:
### Speech Recognition (ASR)
- Multi-language speech recognition systems
- Domain-specific ASR models (agriculture, healthcare, etc.)
- Synthetic vs. natural speech detection
### Text-to-Speech (TTS)
- Multi-language TTS systems
- Domain-adaptive speech synthesis
- Voice quality evaluation (synthetic vs. natural)
### Research Applications
- Cross-domain speech analysis
- Language-specific phonetic studies
- Synthetic speech quality assessment
- Multi-modal AI training
### Commercial Applications
- Voice assistants for specific domains
- Educational pronunciation tools
- Accessibility applications
- Multilingual customer service systems
## Data Quality
- All audio files are accessible via HTTPS URLs with AWS authentication
- Text transcriptions are domain-verified and language-specific
- Unique identifiers ensure data integrity and traceability
- Consistent schema across all entries
- Balanced representation across domains and languages
## License
This dataset is released under the MIT License.
## Citation
```bibtex
@dataset{multi_domain_speech_2025,
title={Multi-Domain Haitian_creole Speech Dataset},
author={Dataset Creator},
year={2025},
languages={ha},
domains={healthcare},
url={https://huggingface.co/datasets/jsbeaudry/med-cre}
}
```
## Acknowledgments
Special thanks to contributors who provided audio recordings and transcriptions across multiple languages and domains to make this comprehensive dataset possible.