Resource: Gram Vanni data set
|Reference||Gram Vanni data set|
|Date of Submission||May 15, 2019, 1:04 p.m.|
|Resource Type||Primary Text|
The Gram Vanni data set consists of 130 hours (21,000 different audio recordings) recorded by 4,000 unique Hindi speakers from the states of Bihar, Jharkhand, and Madhya Pradesh in India (20-25% female, 60% people under 30 years of age, mostly rural).
The data set was collected via a voice-based community media platform that runs over IVR (Interactive Voice Response) telephone systems. Users can call into the system and listen to audio messages, or record their own message in response to messages they hear. This therefore serves as a discussion forum on voice, but without needing the Internet, and suitable even for less-literate populations. The platform is used for discussions on local policies, local news, questions and answers on agriculture, health and social norms, and poetry. All content recorded by the users is manually reviewed before it can be heard by other users over the IVR, to reject content with poor audio quality or editorial violations such as hate speech or false information. The environment for recordings is mostly outdoor, with a medium level of background noise from roadside and public places. Speech samples are stored as sequences of 8 kHz in MP3 files.
An orthographic transcription is provided (transliteration in Latin characters), including the following tagged named entities :