snakers4/open_stt

Comprehensive Russian speech recognition dataset with ~16M utterances spanning 20,000 hours across diverse domains (radio, public speech, audiobooks, YouTube, phone calls) with multiple annotation methodologies including forced alignment, subtitles, and manual annotation. Data is distributed in compressed OPUS format (356GB) alongside WAV archives, with helper utilities for format conversion and on-disk database construction. Designed for training end-to-end ASR systems, with quality-stratified subsets ranging from 70-99% accuracy for different use cases.

818 stars. No commits in the last 6 months.