Sanmati
Collection
5 items • Updated
Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These are spoken answers to agriculture, career, and health prompts. Access is granted by hand. Use is under the KPL BY-NC-SA-FS 1.0 license: noncommercial use only, with attribution, and share-alike for adaptations and for software that incorporates the material.
Log in or Sign Up to review the conditions and access this dataset content.
Sanmati task 4
Spoken answers to agriculture, career, and health prompts in six language folders: bh, bn, hi, ml, mr, and te. All six are in this dataset, including Bhojpuri and Hindi.
Each config loads one language. load_dataset(repo, "bn") reads only the Bengali parquet files.
Audio and text
Audio is not stored inside the parquet. Each parquet shard has a sibling uncompressed tar with the same name:
data/{lang}/train-{iiii}-of-{nnnn}.parquetdata/{lang}/train-{iiii}-of-{nnnn}.tar
A row's audio_in_tar is the wav member name inside tar_file. The wav is the original uncompressed file. text is the transcription of that answer. question is the prompt that was asked. duration is the value from the source export, in milliseconds. speech_present is the source flag for whether speech was detected.
| column | meaning |
|---|---|
| text | transcription of the answer |
| question | prompt that was asked |
| prompt_id | prompt stem, such as agri_q1 or health_scenario_q2 |
| domain | agri, career, or health |
| scenario | scenario text for scenario prompts; empty otherwise |
| scenario_prompt | scenario prompt for scenario prompts; empty otherwise |
| tar_file | path of the sibling tar in this repo |
| audio_in_tar | wav member name |
| sha256 | sha256 of the wav bytes |
| num_bytes | wav size |
| lang | folder code |
| worker_id | worker id from the export |
| datapoint_id | datapoint id from the export |
| duration | source duration, in milliseconds |
| speech_present | source flag |
The Hub audio player does not play a file that lives inside a tar. Read the parquet, then open that member from the matching tar.
Languages
Bhojpuri (bh) is included. Its audio and questions are present. Its transcriptions are not: text is empty on every Bhojpuri row. Hindi (hi) is included, with transcriptions on most rows.
Hours below are duration converted from milliseconds. Sizes are the uncompressed wav bytes.
| lang | name | in dataset | clips | hours | GB | clips with transcription |
|---|---|---|---|---|---|---|
| bh | Bhojpuri | yes | 1,805 | 8.7 | 1.0 | 0 |
| bn | Bengali | yes | 25,154 | 116.6 | 13.4 | 23,819 |
| hi | Hindi | yes | 41,868 | 181.6 | 20.9 | 33,424 |
| ml | Malayalam | yes | 3,423 | 26.8 | 3.1 | 3,354 |
| mr | Marathi | yes | 31,324 | 143.2 | 16.5 | 26,654 |
| te | Telugu | yes | 15,524 | 83.7 | 9.6 | 15,225 |
| 119,098 | 560.6 | 64.6 | 102,476 |
Empty recordings were left out (bn 11, ml 1, mr 4, te 1).
Split
Every row is train. There is no held-out eval split.
- Downloads last month
- 38