IbrahimDayax/somali-asr-synthetic-youtube

Somali ASR Synthetic YouTube Dataset

A Somali-language speech dataset derived from YouTube audio, intended for training and evaluating automatic speech recognition (ASR) and speech-to-text (STT) models. Transcriptions were generated synthetically (silver-standard) via ASR bootstrapping.

Dataset Summary

Split	Samples
train	~4,393
validation	200
test	100
Total	~4,693

• Language: Somali (so)
• Audio format: WAV, 16 kHz, mono, 16-bit PCM
• Total duration: ~28 hours
• Transcript length: 9–522 words per sample (median ~47 words)
• Source: YouTube videos in Somali

Dataset Structure

from datasets import load_dataset

ds = load_dataset("IbrahimDayax/somali-asr-synthetic-youtube")

# Each example contains:
# {
#   "audio":         {"array": [...], "sampling_rate": 16000},
#   "transcription": "burcadada soomaali ah ayaa ...",
#   "language":      "so"
# }

Fields

Field	Type	Description
audio	Audio (16 kHz)	Raw waveform + sampling rate
transcription	string	Somali transcript (silver-standard)
language	string	ISO 639-1 language code (so)

Data Collection

Audio was sourced from publicly available Somali-language YouTube videos. Transcriptions are synthetic (silver-standard), generated via an ASR bootstrap pipeline and have not been manually verified at scale. Quality varies by recording condition and speaker.

Data Cleaning

The published version applies the following fixes over the raw scraped data:

1. Deduplication — 158 rows with duplicate transcription text were removed (69 unique duplicate texts), leaving 4,693 distinct samples.
2. Proper splits — The original dataset had a train/validation data leak (all 200 val samples also appeared in train). Splits are now drawn from disjoint sets using a fixed random seed (42).

Intended Use

• Fine-tuning ASR/STT models for Somali (e.g., Whisper, wav2vec 2.0, MMS)
• Benchmarking low-resource speech recognition
• Speech data augmentation

Limitations

• Transcriptions are silver-standard (machine-generated) and may contain errors.
• Audio quality varies (background noise, compression artifacts from YouTube encoding).
• No speaker metadata (age, gender, dialect) is available.
• Domain is primarily news, political commentary, and social media content.

License

CC BY 4.0 — You are free to share and adapt the material for any purpose, provided appropriate credit is given.

Citation

If you use this dataset, please cite:

@dataset{dayax2024somali_asr,
  author    = {Ibrahim Dayax},
  title     = {Somali ASR Synthetic YouTube Dataset},
  year      = {2024},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/IbrahimDayax/somali-asr-synthetic-youtube}
}