Imagine navigating a market where people communicate in 22 official languages and more than 19,000 dialects. Is it practical to rely solely on a text-based AI chatbot with limited language capabilities?
Facing this challenge head-on, the Indian AI venture Sarvam unveiled its innovative solutions on Tuesday, including a voice-operated AI bot capable of understanding over ten Indian languages. This move anticipates the preference of Indian users to interact in their native languages through voice rather than text. Additionally, Sarvam is introducing a compact language model, a specialized AI tool for legal professionals, and an auditory language model.
“The convenience of conversing in one’s mother tongue is undeniable, especially when typing in Indian languages remains a hurdle,” mentioned Vivek Raghavan, Sarvam AI’s co-founder, during an interview with TechCrunch.
Targeting businesses and corporations, the Bengaluru-based startup is marketing its voice-activated AI bots for various sectors, with an emphasis on customer service. For instance, Sri Mandir, a client that provides religious content, has successfully utilized Sarvam’s AI agent for facilitating payments, recording over 270,000 transactions so far.
Sarvam mentioned that its AI voice agents can be integrated into WhatsApp, applications, and even be used for traditional voice calls.
Securing backing from Peak XV and Lightspeed, Sarvam is setting its AI agent pricing at ₹1 (about 1 cent) per usage minute.
Leveraging a foundational, compact language model known as Sarvam 2B, which is trained on a data set of 4 trillion tokens entirely composed of synthetic data, the startup is making strides in AI development, according to Raghavan.
Despite the caution advised by AI specialists on the use of synthetic data—fabricated by large language models to mimic real-world information—for training AI systems, due to potential inaccuracies, Sarvam has deployed models to refine and advance the initial synthetic data sets used for training.
Raghavan highlighted that Sarvam 2B comes at a fraction of the cost of similar offerings in the market and the firm plans to make the model open source for further community development.
“While foundational large language models are intriguing, smaller language models can provide a more tailored, cost-effective, and quicker experience for specific needs,” stated Raghavan. “For occasional inquiries, large models are adequate, but for daily interactions in the millions, smaller models are the way to go.”
Furthermore, Sarvam is introducing Shuka, an auditory-language model built atop its Saaras v1 audio decoder and Meta’s Llama3-8B Instruct, which will also be open-sourced. This enables developers to craft voice interfaces using the startup’s translation, TTS, and additional modules.
Another notable innovation is “A1,” a generative AI workbench tailored for lawyers to assist with regulatory research, document drafting, redaction, and data extraction.
Sarvam stands among the select Indian startups advocating for AI applications that meet the nation’s specific needs and support the government’s ambition to develop a customized AI infrastructure.
Nations worldwide are increasingly embracing “sovereign AI” – the pursuit of AI infrastructure managed and developed within a country. Such initiatives aim to protect data privacy, boost economic growth, and customize AI to fit cultural idiosyncrasies. With significant investments from the United States and China, India is keeping pace with its “IndiaAI” program, emphasizing language-specific models.
Under the IndiaAI initiative, one project, the IndiaAI Compute Capacity, aims to set up a supercomputer fueled by at least 10,000 GPUs. Another, named Bhashini, seeks to broaden the accessibility of digital services across a variety of Indian languages.
Raghavan expressed Sarvam’s readiness to support the IndiaAI initiative, stating the company’s eagerness to collaborate with the government if the chance arises.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
