Krishna Permi's journey began with a simple desire to help his son with Kannada homework. This seemingly mundane task led to the creation of Akshar, a free keyboard app that supports 21 Indian languages. Permi's story highlights the intricate relationship between technology, privacy, and cultural identity, particularly in the context of Indian languages.
Permi's initial challenge was to type in Kannada on his iPhone, a task that required an internet connection and a web tool like Google Input Tools. This experience sparked his determination to build a privacy-preserving keyboard app. The key issue was the data transmission to an external server, which prompted Permi to seek a fully private, offline solution.
The quest for privacy led Permi to AI4Bharat, a research initiative focused on natural language processing for Indian languages. They provided an open-source transliteration model, IndicXlit, which converts phonetic English input into native Indian scripts. This model, trained across 21 Indian languages, became the foundation for Akshar.
However, integrating IndicXlit into a keyboard presented challenges. The iOS keyboard extensions have strict memory limits, and the model's size needed to be reduced without compromising accuracy. Permi employed a technique called quantisation, reducing the numerical precision of the model's weights to fit it within the keyboard's memory constraints.
Akshar's unique features include its offline capability, lack of user data collection, and support for 21 Indian languages, including lesser-known ones like Konkani, Bodo, Kashmiri, and Manipuri. This level of inclusivity is a direct response to the technology industry's tendency to prioritize languages with larger user bases and business viability.
Despite its impressive features, Akshar is still a work in progress. Permi acknowledges the need for improvements in accuracy, especially with numbers and autocorrect. He is actively working on these aspects to enhance the user experience.
The app's reception has been positive, with users finding it useful for various purposes, from sending Marathi WhatsApp messages to second-generation Indians in the United States who grew up typing on English keyboards. The app's free nature, supported by local processing via Core ML, ensures no server costs or API fees.
Looking ahead, Permi plans to bring Akshar to Android users, considering the majority of Indian smartphone users on that platform. He also aims to integrate new language models as they become available, with Gondi, an officially recognized Indian language, as a potential future addition.
Permi's story underscores the potential of technology to empower individuals and communities, especially in the realm of language. His commitment to privacy, accuracy, and inclusivity sets a precedent for developers to create tools that respect user data and cater to diverse linguistic needs.
