According to cofounder Vivek Raghavan, Sarvam AI, an artificial intelligence (AI) startup founded to develop fundamental models for the Indian language, is getting ready to release its first voice-to-voice endpoint solution for commercial use. This tool will assist businesses with voice-related tasks like customer care.
“At the SaaSBoomi annual event in Chennai on March 7, Raghavan addressed SaaS founders and said, ‘You can expect some voice-to-voice endpoints with at least 10 Indian languages and you can expect some experiences built on this and also example experience for people to use and build on top of it.”
He also mentioned that startups and companies wishing to create or incorporate voice experiences into their offerings might make use of this platform, particularly for Indian languages.
The company has launched OpenHathi-Hi-v0.1, the OpenHathi series’ first large language model (LLM) in Hindi. Based on Meta AI’s Llama2-7B architecture, Sarvam AI claims that the model performs comparably to GPT-3.5 for Indic languages.
In comparison to GPT-4 and GPT-3, Raghavan stated during his speech at SaaSBoomi that OpenHathi has demonstrated better performance in the translation of English to Hindi.
Raghavan also discussed the difficulties the company is now having gathering data and the costs associated with tokens as they work to construct the basic model for the Indian language.
“Quality data collecting presents hurdles, and tokenization costs are higher for Indian languages. Speaking at the SaaSBoomi event, he continued, “There are also evaluation issues for something fresh like what we are doing.
In order to network and share knowledge, a casual group of SaaS creators called SaaSBoomi was established in 2015. Currently, the area is home to around 800 businesses.
On March 6, the AI4Bharat research lab at IIT Madras unveiled IndicVoices, an open-source natural and voice collection that encompasses 22 Indian languages.
The goal of this dataset, according to a blog post by AI4Bharat, was to gather impromptu speech in Indian languages. Bhashini, supported by the Ministry of Electronics and Information Technology, the Ekstep Foundation, and Nilekani Philanthropies, provides funding for IndicVoices.
“The ecosystem will benefit greatly from this. Startups will find high-quality data from AI4 Bharat. Nevertheless, more work needs to be done in order to create a foundational model, according to Raghavan.
In December 2023, Sarvam became the first AI firm in India to raise $41 million during its Series A fundraising round, which was headed by Lightspeed Ventures and included participation from Peak XV Partners and Khosla Ventures.
Founded in July 2023, Sarvam is a full-stack Generative AI provider, encompassing research-driven innovations in training custom AI models and an enterprise-grade platform for authoring and deployment. Pratyush Kumar, a former employee of Infosys co-founder Nandan Nilekani-backed AI4Bharat, and Vivek Raghavan founded the company.
Technologist and entrepreneur Raghavan, who played a key role in creating Digital Public Goods (DPGs) like Aadhaar, said that Sarvam will collaborate with Indian businesses to co-develop AI models that are domain-specific using their data.
