Skip to content

Dataset · bond005

SberDevices Golos Russian speech corpus subset (crowd)

bond005/sberdevices_golos_10h_crowd

16 kHz Russian speech audio recordings with manual annotations from the crowdsourced part of the Golos corpus for ASR tasks.

What's inside

The repository features a subset (10,000 to 100,000 examples) of the well-known Russian speech corpus Golos, prepared by the SberDevices team. The full corpus contains about 1,200 hours of audio (16 kHz) from crowdsourcing and smart device interactions, manually annotated with text and split into train and test.

What you can do with it

  • Fine-tuning Russian speech recognition (ASR) models such as Whisper or Conformer
  • Evaluating transcription quality on clean crowdsourced speech
Tasks
Automatic speech recognition, Audio classification
Languages
Russian
Size
10K–100K rows
Downloads this month
25.8K
Likes
6

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("bond005/sberdevices_golos_10h_crowd")