datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
nirantar
Nirantar
Nirantar speech dataset (22 Indian languages) in Hugging Face format. Source: AI4Bharat/Nirantar.
Downloading language-wise subsets
Each language is a separate configuration (subset), so you can load only one language (like ai4bharat/Rasa):
from datasets import load_dataset, get_dataset_config_names
# Single language (only that language's parquet is downloaded)
ds = load_dataset("adjaysagar/nirantar", "hi", trust_remote_code=True) # Hindi
ds["train"] #… See the full description on the dataset page: https://huggingface.co/datasets/shiprocket-ai/nirantar.vaani-non-null
Vaani by Language — Non-Null Transcript Subset
Processed output from ARTPARK-IISc/Vaani: transcript-not-null filtered, reorganized by language. Hours computed from WAV duration in each parquet (run compute_vaani_hours.py to regenerate).
Non-null hours per language
Language
Hours
Hindi
963.01
Bengali
154.39
Telugu
151.49
Kannada
151.42
Marathi
68.46
Odia
52.48
Chakma
49.81
Garo
47.33
Malayalam
41.90
Nepali
38.17
Assamese
30.35… See the full description on the dataset page: https://huggingface.co/datasets/shiprocket-ai/vaani-non-null.