shraavb/spanish-slang-stt-data

Spanish Regional Speech-to-Text Dataset

A multilingual Spanish speech recognition dataset covering 4 regional dialects for fine-tuning Whisper and other ASR models.

Dataset Description

This dataset contains ~39,000 audio samples with transcriptions across 4 Spanish-speaking regions:

Region	Samples	Description
Mexico	17,725	Mexican Spanish including CIEMPIESS corpus
Spain	11,360	Castilian Spanish from TEDx and Common Voice
Argentina	5,839	Rioplatense Spanish from OpenSLR
Chile	4,374	Chilean Spanish from OpenSLR

Intended Use

• Fine-tuning Whisper for regional Spanish dialects
• Training regional speech recognition models
• Studying dialectal variations in Spanish speech

Dataset Structure

Data Fields

Field	Type	Description
audio_path	string	Relative path to audio file
transcript	string	Text transcription
duration	float	Audio duration in seconds
language	string	Language code (es)
region	string	Geographic region (mexico, spain, argentina, chile)
dataset	string	Source dataset name
original_path	string	Original file path
dialect_markers	list	Regional dialect indicators (optional)
speaker_id	string	Speaker identifier (where available)

Data Splits

Split	Samples
Train	31,438
Validation	3,929
Test	3,931

How to Use

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("shraavb/spanish-slang-stt-data")

# Access training data
train_data = dataset["train"]

# Filter by region
mexico_samples = [s for s in train_data if s["region"] == "mexico"]

# Example sample
print(train_data[0])
# {
#   'audio_path': 'data/processed/mexico/ciempiess/sample.wav',
#   'transcript': 'Hola, buenos días',
#   'duration': 3.5,
#   'language': 'es',
#   'region': 'mexico',
#   ...
# }

Filter by Region

from datasets import load_dataset

dataset = load_dataset("shraavb/spanish-slang-stt-data")

# Get only Spanish (Castilian) samples
spain_train = dataset["train"].filter(lambda x: x["region"] == "spain")
print(f"Spain samples: {len(spain_train)}")

Source Datasets

Dataset	Region	License
CIEMPIESS	Mexico	CC BY-NC-SA 4.0
TEDx Spanish	Spain	CC BY-NC-ND 4.0
Common Voice	Spain, Mexico, Argentina	CC0
OpenSLR Argentine	Argentina	CC BY 4.0
OpenSLR Chilean	Chile	CC BY 4.0

Audio Specifications

• Sample Rate: 16,000 Hz
• Format: WAV (mono)
• Duration Range: 0.5s - 30s (filtered for Whisper compatibility)

Training Scripts

See spanish-slang-stt for:

• Fine-tuning scripts for Whisper
• Regional model training (Approach 2)
• Evaluation scripts

Citation

@misc{spanish-slang-stt-2024,
  title={Spanish Regional Speech-to-Text Dataset},
  author={Shraavasti Bhat},
  year={2024},
  url={https://huggingface.co/datasets/shraavb/spanish-slang-stt-data}
}

License

This dataset is released under CC-BY-4.0. Individual source datasets may have additional restrictions (see Source Datasets table).