In 2019, Reliance Industries acquired Indian language technology company Reverie Language Technologies for around ₹190 crore, or about $27 million. Reverie was founded in 2009 by Vivekanand Pani, his brother Arvind Pani and S K Mohanty to address the challenges of using Indian languages on phones and digital platforms.
The acquisition came after Reverie had raised just over $4 million. For founder Vivekanand Pani, however, the deal was not the end of his work in Indian language technology.
After nearly three decades in the field, Pani has started a new company called Coremantle. The company is focused on a major challenge facing artificial intelligence today — the lack of natural and useful data needed to help AI systems understand how Indians actually speak.
The challenge is large because India has 22 official languages, many other major languages and thousands of dialects. Differences in accents, regional speech, code-switching and local vocabulary make it difficult to build AI systems that can work accurately across the country.
Pani believes simply creating more data, especially synthetic data, may not solve the problem. He says AI companies need more natural conversations that reflect how people speak in real life.
Speech recognition systems can perform very well in controlled tests, but their accuracy can fall sharply during everyday conversations. English has an advantage because it has decades of digital data, while many Indian languages have much less digital content.
Indian Languages Have Many Variations
Pani points to Odia, his native language, as an example. Although Odia is already a relatively low-resource language online, there are many different ways people speak it.
Regional dialects, differences between formal and everyday speech, and the use of Hindi and English can all change how people communicate.
This means collecting large amounts of data is not enough. The data also needs to represent the way people naturally speak.
Coremantle uses 2,000 hours of usable data for one language as an example of the scale needed, but Pani says the bigger challenge is collecting such data naturally.
Why Synthetic Data May Not Be Enough
With limited Indian-language data available, many AI companies are turning to synthetic data. This can include AI-generated content, translated material or scripted recordings.
Pani argues that such data cannot fully capture natural conversations. People often interrupt themselves, change words, mix languages, shorten sentences and use local expressions while speaking.
As a result, a speech system may perform well with scripted recordings but struggle when used in everyday conversations.
Building Better Language Data
Pani also sees the current problem as the result of years of limited digital content in Indian languages.
India has worked on Indian-language computing since the 1980s, but much of the internet later developed around English. This left many Indian languages with far less digital data compared with English.
Pani started Reverie in 2009 when Indian-language technology was still seen as a difficult business opportunity. The company remained largely bootstrapped during its early years before eventually being acquired by Reliance Industries.
With Coremantle, Pani is now trying to address the data gap that continues to affect Indian-language AI.
The company is still in its early stages and is currently testing its approach. Its focus is not simply on collecting more data, but on collecting data that better represents real human conversations.
For India’s growing AI industry, the challenge may therefore be not just how much Indian-language data is available, but how closely that data reflects the way people actually speak.
