karya/sanmati-leading-question

Sanmati Leading Questions

Sanmati Leading Questions contains spoken answers to leading questions in six Indian languages. A worker is asked a prompt, and the recording is their spoken response. This is not a story-recitation dataset.

text is the worker's spoken answer. question is the prompt they were asked. Simple <b> markup around a prompt is removed before the prompt is stored. variant is main or v3. Bengali, Hindi, Malayalam, Marathi, and Telugu include both variants. Bhojpuri has main only.

What is in each row

Column	Description
text	Spoken answer, from acs_transcription_recordingN.
question	Prompt the worker was asked, from questionN, with simple HTML bold tags removed.
question_index	1-based index N of that prompt on the worker's export row.
tar_file	Repository path of the uncompressed tar shard that holds the WAV.
audio_in_tar	Member name of the WAV inside that tar.
sha256	SHA-256 hex digest of the WAV bytes.
num_bytes	WAV size in bytes.
lang	Language code: bh, bn, hi, ml, mr, or te.
variant	main or v3.
worker_id	Worker identifier from the export.
duration	durationN from the export, in milliseconds.

Audio files are stored in uncompressed tar shards next to the parquet metadata:

• data/{lang}/train-00000-of-NNNNN.parquet
• data/{lang}/train-00000-of-NNNNN.tar

Each parquet row points at one WAV through tar_file and audio_in_tar. Rows from main and v3 share a language's shard sequence. If the same WAV basename occurs in both variants, both rows are kept and the later tar member is stored as {variant}/{basename}.

Hugging Face's audio player may not play members packed inside tar files. Open the sibling tar and read audio_in_tar.

Languages

Choose a configuration to load one language. Each configuration has a single train split. There is no held-out evaluation split.

Config	Language	Examples	Variants
bh	Bhojpuri	0	0 main
bn	Bengali	90,353	32,653 main / 57,700 v3
hi	Hindi	58,940	14,545 main / 44,395 v3
ml	Malayalam	84,506	53,507 main / 30,999 v3
mr	Marathi	83,640	28,228 main / 55,412 v3
te	Telugu	67,279	23,110 main / 44,169 v3

There are 384,718 usable clips in total. Usable clips are WAV files that join to a non-empty transcription. Bhojpuri currently has no usable clips because every BH WAV in this export lacks a transcription.

Loading the data

from datasets import load_dataset

bengali = load_dataset("karya/sanmati-leading-question", "bn", split="train")
print(bengali[0]["question"])
print(bengali[0]["text"])
print(bengali[0]["tar_file"], bengali[0]["audio_in_tar"])

Access

Access is manually approved. Accept the conditions on the dataset page before downloading files.

Acknowledgments

We gratefully acknowledge the communities and individuals who contributed these recordings and transcripts.

License

This work is licensed under a KPL BY-NC-SA-FS 1.0 license. This license allows reusers to distribute, remix, adapt, build upon, and incorporate into software systems, the material in any medium or format for noncommercial purposes only, and only so long as attribution is given to the creator. If you remix, adapt, build upon, or incorporate into software systems, you must license the modified material, including material generated by the software system, under identical terms, and license the software system under the GNU General Public License.

See LICENSE in this repository and the full text at karya-inc/kpl.

Commercial Use

If you are interested in using this dataset for commercial purposes, please contact us. Profits from commercial licensing will be distributed as royalties to the community members who contributed to this dataset.

Contact

For questions, issues, or contributions, open an issue on the dataset repository or contact us directly.