Evanroubert/Armenian_audio

Armenian 7,500-Hour Raw Speech Dataset (Public Gated)

A large-scale collection of approximately 7,500 hours of Armenian speech, prepared for modern speech AI research and large-scale model training.

The audio has been processed using Voice Activity Detection (VAD), converted to 16 kHz mono PCM, and packaged into Snappy-compressed Parquet shards (~500 MB each) for efficient streaming and seamless integration with the Hugging Face datasets library.

Designed for training and pre-training state-of-the-art speech models, including HuBERT, Wav2Vec 2.0, WavLM, Whisper, and other ASR, speech representation, and speech foundation models.

Features

• ~7,500 hours of Armenian speech
• Voice Activity Detection (VAD) applied
• 16 kHz mono PCM audio
• Snappy-compressed Parquet shards (~500 MB each)
• Optimized for high-performance data loading
• Fully compatible with the Hugging Face datasets library

Access

Public Gated Dataset. Access requires manual approval from the dataset owner.