Datasets:
Multilingual speech for GigaAM
Verified release: 6e7a2e5bd487ac44f6070af682d2f0ca033a572042f32cdde9c2a7afd1304b64. Sources: common_voice, fleurs_en, fleurs_ru, fleurs_uz, librispeech_en_clean, omnilingual_uz, russian_librispeech, sova_rudevices_ru, usc, uzbekvoice, youtube_it, youtube_news, youtube_podcasts.
Each row contains audio (embedded mono 16 kHz FLAC), duration in seconds, text and its identical alias transcription, plus dataset_name (the source ID), with source, language/accent evidence, speaker/session IDs and provenance retained for auditing. Official dev/test roles remain separate. Rustam's spontaneous benchmark is excluded and must remain test-only.
For the official GigaAM trainer, materialize the audio to local files and use TSV columns path, duration, transcription. The original GigaAM feature extractor computes features at training time.
This collection preserves per-row upstream license information; it does not grant a new blanket license. Consult the linked source revisions before redistribution. Automated checks do not certify transcript accuracy or detect every near-duplicate.
Full report and charts · Source revisions
| split | clips | hours |
|---|---|---|
| train | 818837 | 1,285.365 |
| validation | 28015 | 39.633 |
| test | 53406 | 70.736 |
- Downloads last month
- 36