Speech Recognition
AI technologies for speech recognition hold promise to enable a variety of use-cases, including providing input tools for lower literacy users. AI4Bharat has a deep focus on building state-of-the-art speech recognition in many Indian languages and in ensuring that the models are deployable on mobile devices.
Datasets
Dhwani Dataset
17,000 hours of raw speech data for 40 Indian languages from a wide variety of domains including education, news, technology, and finance
IndicSUPERB
A benchmark of speech recognition tasks including ASR, speaker verification, speaker identification, language identification, query by example, and keyword detection for 12 Indian languages.
Shrutilipi
Over 6,400 hours of labelled audio across 12 Indian languages mined and aligned from audio broadcasts and PDF transcripts from All India Radio.
Svarah
Over 6,400 hours of labelled audio across 12 Indian languages mined and aligned from audio broadcasts and PDF transcripts from All India Radio.
Models
On-Device ASR
Much smaller ASR models which can be quantized and executed on Android devices to support privacy-preserving inference on personal devices.
IndicWav2Vec
State-of-the-art open-source ASR models for 9 languages (including Nepali and Sinhala) as measured on public benchmarks.
Tools
Chitralekha
Over 6,400 hours of labelled audio across 12 Indian languages mined and aligned from audio broadcasts and PDF transcripts from All India Radio.
Our Partners
DesiCrew
We are working with Desicrew to collect voice samples from across 500 districts in the country. We are beginning with TamilNadu in August 2022.
Karya
Karya is an application developed at Microsoft Research and maintained by Karya Inc. We are using Karya to collect voice samples from across the country.
NPTEL
We are working with NPTEL to deploy the Chitralekha tool for subtitling and translating higher education videos.