Google I/O Regional Connect: Tech giant open-sources Indian speech data, location info

Alphabet-owned tech giant Google has open-sourced research models and datasets to help developers build solutions with India-focused speech data and location information, Manish Gupta, head of Google Research India told ET.

Through its collaboration with Google as part of Project Vaani, Bengaluru's research university Indian Institute of Science is now open sourcing the first set of speech data comprising over 4,000 hours across 38 languages.

"Out of the 125 languages we committed to supporting, 75 had a zero data corpus which means the amount of data corpus available in digital form to AI researchers was zero. In the 4,000 hours of speech data that has been put out, for a few languages, it's the first such instance that digital data has been made available. We could expect innovations in these zero corpus languages now," Gupta said speaking on the challenges arising out of the diversity of Indian languages.


The company for the first time hosted Google I/O connect for developers in India in Bengaluru on Wednesday.

Senior Google executives including Ambarish Kenghe - Vice President, Product, Google Pay; Rahul Sukthankar - VP, Google Research; Will Grannis - VP & CTO, Google Cloud; Mathew McCollough - VP, Product, Android Developer; and Una Kravets, Developer Relations Engineer, counted among the speakers.

Google also unveiled a slate of AI tools and technologies for India to support innovation among local developers.

While noting that there are already more than 60 generative AI startups in India, Kenghe said, to help developers build AI-powered products, Google is making its large language model accessible through PaLM API, MakerSuite, and features on Vertex AI.

Google is also releasing Open Buildings information of over 200 million buildings in the country to help organisations plan infrastructure projects.

"Google already offers Plus Codes for addresses. If one wants to develop delivery-related apps, with addresses, one can take advantage of this location information, which is available openly. This is an AI system which is able to identify the footprint of a building at a city or district scale based on satellite imagery. It is difficult for a town planner to identify the density of a bunch of smaller buildings, and this addresses it," Gupta said.

It could be really useful for town planning and disaster management and for people building apps around that, he said.

The company will also soon be rolling out a Trusted Tester programme for developers to access its healthcare artificial intelligence model application programming interfaces that can identify medicine names within handwritten prescriptions.

"This is a hard problem to solve. This capability is now available if you're a startup or a developer under the tester programme. An e-pharmacy can try this out for example. We want to get pilots done and developers give us feedback. With the help of our partners, we can get it into the hands of users," Gupta said.

Google Cloud is launching an accelerator programme for the government-owned e-commerce marketplace the Open Network for Digital Commerce (ONDC) to help India’s digital sellers build and scale their digital commerce operations, Grannis said.

As part of this initiative, the company is open sourcing a ready implementation of ONDC infrastructure and core APIs to facilitate scalability and security, and enabling access to its Retail AI technology and PaLM API.

Google Cloud is also introducing a startup credits programme where organisations that enable ONDC are eligible to apply for a US$25,000 grant.

Sukthankar announced Google is also open-sourcing the SeeGull Database, a global stereotype benchmark with broad geo-cultural coverage including stereotypes existing within India, to evaluate and mitigate biases in Natural Language Processing.

In addition, Google Maps Platform launched Address Descriptors – an India-first experimental feature available in 25 Indian cities – to make it easier for customers to find and communicate addresses using relevant landmarks and area names.

Google also shared how it is helping developers build for the web, and announced new features to enable developers to build engaging mobile and multi-device experiences.

Overwhelming percentage of data that large language models have been trained on are mature languages like English, French and German etc, so ensuring their capabilities for reasoning, generating content fluidly, answering questions etc are carried on to new resource languages is a significant challenge, Gupta said.