cryptolock/common-voice-22-en-teens

Common Voice 22 - English Teens Subset

This dataset is a filtered subset of Mozilla's Common Voice Corpus 22.0, containing only recordings from teenage speakers in English.

Dataset Structure

Data Instances

Each instance contains:

• audio: Audio file with embedded bytes, sampling rate, and path
• text: Transcription of the spoken text

Data Splits

This dataset provides a single unified split containing all 77,500 samples. Users can create their own train/validation/test splits based on their specific needs.

Split	Samples
train	77,500

Speaker Demographics

• Unique Speakers: 2,166 teenage contributors
• Average Recordings per Speaker: ~35.9

Gender Distribution:

• Male/Masculine: 41,539 (53.4%)
• Female/Feminine: 25,282 (32.5%)
• Unspecified: ~10,974 (14.1%)
• Other: 12 (0.01%)

Audio Specifications

• Sample Rate: 16,000 Hz (16 kHz)
• Channels: Mono
• Format: MP3
• Audio Storage: Embedded directly in the dataset (no external files needed)

Text Transcriptions

• Total Transcriptions: 77,500
• Average Length: ~62 characters
• Range: 3 to 17,096 characters

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("cryptolock/common-voice-22-en-teens")

# Access the data
train_data = dataset['train']

# Example: Iterate through samples
for sample in train_data:
    print(f"Text: {sample['text']}")
    print(f"Audio sampling rate: {sample['audio']['sampling_rate']}")
    # Access audio array: sample['audio']['array']

Creating Custom Splits

from datasets import load_dataset

# Load dataset
dataset = load_dataset("cryptolock/common-voice-22-en-teens")

# Create train/val/test splits (e.g., 80/10/10)
train_test = dataset['train'].train_test_split(test_size=0.2, seed=42)
test_val = train_test['test'].train_test_split(test_size=0.5, seed=42)

splits = {
    'train': train_test['train'],      # 80%
    'validation': test_val['train'],   # 10%
    'test': test_val['test']           # 10%
}

print(f"Train: {len(splits['train'])} samples")
print(f"Validation: {len(splits['validation'])} samples")
print(f"Test: {len(splits['test'])} samples")

Fine-tuning Whisper Example

from datasets import load_dataset
from transformers import WhisperProcessor, WhisperForConditionalGeneration

# Load dataset and model
dataset = load_dataset("cryptolock/common-voice-22-en-teens")
processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")

# Prepare data
def prepare_dataset(batch):
    audio = batch["audio"]
    batch["input_features"] = processor(
        audio["array"],
        sampling_rate=audio["sampling_rate"],
        return_tensors="pt"
    ).input_features[0]
    batch["labels"] = processor.tokenizer(batch["text"]).input_ids
    return batch

# Process dataset
dataset = dataset.map(prepare_dataset, remove_columns=dataset.column_names)

# Continue with training...

Use Cases

This dataset is particularly useful for:

• Training and fine-tuning speech recognition models on teenage voices
• Studying acoustic characteristics of adolescent speech
• Building age-specific ASR systems optimized for youth demographics
• Voice activity detection for teenage speakers
• Educational and research purposes in speech processing
• Benchmarking ASR models on teen voice data

Dataset Creation

Source Data

This dataset was derived from the Mozilla Common Voice Corpus 22.0 released on June 20, 2025.

Data Collection Process

1. Downloaded the full Common Voice 22.0 English dataset
2. Filtered all recordings where the speaker's self-reported age group was "teens"
3. Extracted corresponding audio clips and metadata
4. Embedded audio bytes into Arrow format for self-contained distribution
5. Combined into a single unified dataset (no pre-defined splits)

Annotations

The transcriptions were crowd-sourced and validated by Mozilla Common Voice contributors. Each recording includes:

• Human-verified text transcription
• Speaker metadata (age group, gender)
• Quality metrics (upvotes/downvotes from validators)

Quality

• All samples verified as teen age group
• Audio validated and embedded
• Self-contained dataset (no external file dependencies)
• Ready for immediate use

Licensing

This dataset inherits the CC0 1.0 Universal (Public Domain) license from Mozilla Common Voice.

You are free to:

• ✅ Use the dataset for any purpose (commercial or non-commercial)
• ✅ Redistribute and modify without restrictions
• ✅ Use without attribution (though attribution is appreciated)

No rights reserved. The dataset is dedicated to the public domain.

Citation

If you use this dataset, please consider citing the original Mozilla Common Voice project:

@inproceedings{commonvoice:2020,
  author = {Ardila, R. and Branson, M. and Davis, K. and Henretty, M. and Kohler, M. and Meyer, J. and Morais, R. and Saunders, L. and Tyers, F. M. and Weber, G.},
  title = {Common Voice: A Massively-Multilingual Speech Corpus},
  booktitle = {Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020)},
  pages = {4211--4215},
  year = {2020}
}

And optionally mention this filtered subset:

@dataset{common_voice_22_teens,
  author = {cryptolock},
  title = {Common Voice 22 - English Teens Subset},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/cryptolock/common-voice-22-en-teens}
}

Additional Information

Dataset Curators

This subset was created and curated by cryptolock to make age-specific speech data more accessible for research and development.

Supported Tasks

• Automatic Speech Recognition (ASR)
• Speech-to-Text
• Voice Activity Detection
• Speaker Verification
• Acoustic Analysis

Languages

• English (en)

Source

• Original Dataset: Mozilla Common Voice 22.0
• Release Date: June 20, 2025
• Subset Created: November 2025

Known Limitations

• Dataset contains 77,500 samples out of the original 77,807 teen samples from CV22 (307 samples or 0.4% missing due to processing)
• Age groups are self-reported by contributors
• Audio quality varies based on recording equipment used by contributors
• Some transcriptions may contain minor errors or variations

Updates

• v1.0 (November 2025): Initial release with 77,500 samples

Contact

For questions about this dataset subset, please open a discussion on the dataset page.

For questions about the original Common Voice dataset, visit https://commonvoice.mozilla.org/

---

Acknowledgments: This dataset would not be possible without the thousands of Mozilla Common Voice contributors who donated their voices and time to create this public domain resource.