Lightspeed-backed homegrown generative artificial intelligence (AI) startup, Sarvam AI is working towards launching its first commercial ‘voice-to-voice endpoint’ in the next six to twelve months, cofounder Vivek Raghavan said on Thursday at a session at an annual event for Software-as-a-Service (SaaS) startups, SaaSBoomi.“You can expect these kinds of voice-to-voice endpoints in at least 10 (Indian) languages and you can expect some experiences built upon this and also there are example experiences in the sense that there are ways for people to build things on top of it,” Raghavan said.Elevate Your Tech Prowess with High-Value Skill CoursesOffering CollegeCourseWebsiteIIM KozhikodeIIMK Advanced Data Science For ManagersVisitIIM LucknowIIML Executive Programme in FinTech, Banking & Applied Risk ManagementVisitIIT DelhiIITD Certificate Programme in Data Science & Machine LearningVisitHe emphasised that the LLMs need to be voice-based LLMs which are also ‘agentic and action-oriented.’ They key thing, he said, is that it needs to work well in colloquial languages.“Building Indic language models is important but that alone probably won’t lead to the result we are looking for which is the widespread use of Gen AI in India,” Raghavan said. “Voice is the primary way by which people will access LLMs in India. We need to have voice-driven interfaces/systems, only then can we have accessibility to a large number of people.”Sarvam AI released its first open-source Hindi language model called OpenHathi-Hi-0.1 in December last year. The firm said the AI model is the first in a series of models which will “make contributions to the ecosystem with open models and datasets to encourage innovation in Indian language AI.”Also read | Ola founder Bhavish Aggarwal’s Krutrim AI turns unicorn with $50 million funding from Matrix, othersDiscover the stories of your interestBlockchain5 StoriesCyber-safety7 StoriesFintech9 StoriesE-comm9 StoriesML8 StoriesEdtech6 StoriesRaghavan stated that OpenHathi has shown improved performance when it comes to English to Hindi translation when compared to GPT-4 and GPT-3.“We looked at a translation task as a proxy to understand how well the understanding of these languages was,” he explained. “GPT’s performance is decent in Hindi but if you go to languages beyond that, you see that the performance really falls away. And if you’re looking at Indic languages, it gives you an idea that when you do these custom things, you can make these models work significantly better.”Speaking of the challenges, Raghavan said that almost all LLMs are trained on English data and that Indic languages are not well represented.Further, the quality of the data is also poor. Tokenization and evaluation were among the other challenges, he said. Sarvam raised $41 million in its Series A funding round led by Lightspeed Ventures with participation from Peak XV Partners and Khosla Ventures in December last year. The company was founded in July 2023 by Vivek Raghavan and Pratyush Kumar, who previously worked at Infosys co-founder Nandan Nilekani-backed AI4Bharat.
