Breaking Barriers in Indic Language AI

In partnership with

Dear Reader,

Flipped.ai’s weekly newsletter is read by more than 75,000 professionals, entrepreneurs, decision-makers, and investors around the world.

In this newsletter, we’re excited to share that India has taken a significant leap in AI with Sarvam AI's launch of Shuka v1, the country's first open-source audio language model built on Meta's Llama. This breakthrough enables voice-based interactions in ten Indian languages, making AI more accessible for businesses, government agencies, and everyday users. In this edition, we dive into how Sarvam AI is transforming multilingual voice AI, its technical innovations, and the future of Indic language AI.

India's First Open-Source Audio Language Model: Sarvam AI's Breakthrough with Llama

Introduction

India is witnessing a revolutionary shift in artificial intelligence (AI) with the development of its first open-source audio language model, Shuka v1, by Bengaluru-based startup Sarvam AI. Built on Meta’s Llama model, this pioneering AI solution supports voice-based interactions in ten Indian languages, bridging the digital divide and enabling businesses and government agencies to deploy AI-powered solutions on-premises.

Sarvam AI’s efforts in training large language models (LLMs) for Indic languages underscore a broader commitment to creating AI systems that cater to India's linguistic and cultural diversity. This article explores Sarvam AI’s groundbreaking innovations, their significance in India's AI landscape, and the challenges the company faces as it strives to establish itself as a leader in generative AI.

The vision behind Sarvam AI

Sarvam AI was founded with the mission of empowering over a billion Indians by creating full-stack generative AI solutions tailored for regional languages. Recognizing that a significant portion of the population prefers voice-based interactions over text, Sarvam AI developed Shuka v1 to improve accessibility and usability of AI systems across different sectors, including customer support, education, and government services.

Shuka v1: India’s First Open-Source Audio Language Model

Leveraging Llama for Indic Voice AI

Shuka v1 is a significant milestone in India’s AI development, marking the first open-source audio language model for Indic languages. The model utilizes Meta’s Llama as the decoder, which processes audio tokens generated by Sarvam AI’s proprietary audio encoder. These tokens capture phonetic and linguistic nuances from spoken input, allowing Shuka to generate accurate text-based responses in multiple Indian languages.

Dr. Pratyush Kumar, co-founder of Sarvam AI, emphasized the importance of using Llama:

Addressing linguistic gaps in AI

One of the major challenges in AI development in India has been the lack of robust voice models for regional languages. Shuka v1 bridges this gap by offering:

• Support for ten Indic languages, including Hindi, Gujarati, Kannada, Marathi, and Tamil.

• Voice-first interactions, catering to users who prefer speaking over typing.

• On-premises deployment, ensuring data privacy and regulatory compliance.

The open-source nature of Shuka v1 allows businesses and government institutions to leverage AI without worrying about data security issues associated with cloud-based models.

Technical innovations in Shuka v1

Architecture and training strategy

Sarvam AI implemented a sophisticated technical framework to develop Shuka v1 efficiently. The core architecture consists of:

• Llama 3’s 8B-Instruct version, selected for its balance of computational efficiency and accuracy.

• A 60M-parameter projector layer, enabling seamless transformation of audio data into Llama-compatible text embeddings.

• A custom-built audio encoder, Saaras v1, designed specifically for Indic languages.

To optimize computational resources, the team adopted a frugal training approach, fine-tuning only the projector layer while keeping Llama and Saaras v1 frozen. This strategy minimized resource consumption while maintaining high linguistic accuracy.

Generating high-quality training data

Since training resources were limited, Sarvam AI used synthetic data generation to create training datasets. The process included:

• Curating high-quality question-answer pairs in Indic languages.

• Processing them through Llama 3 to generate gold-standard answers.

• Fine-tuning the projector layer using this dataset to enhance speech-to-text accuracy.

These techniques ensured that Shuka v1 could efficiently decode spoken queries while maintaining contextual relevance and linguistic precision.

Expanding Beyond Audio: The Launch of Sarvam 1

A specialized LLM for Indian languages

Following the success of Shuka v1, Sarvam AI introduced Sarvam 1, a large language model (LLM) trained specifically for Indian languages. This model is built on a 2-billion-parameter architecture and supports:

• Bengali, Gujarati, Hindi, Marathi, Malayalam, Kannada, Odia, Tamil, Telugu, Punjabi, and English.

• A specialized tokeniser developed by Sarvam AI.

• Training on 4 trillion tokens, leveraging Nvidia’s H100 Tensor Core GPUs.

Sarvam 1 was trained using data and computing resources within India, making it the country’s first sovereign LLM fully developed domestically.

Applications and market impact

Sarvam AI’s LLMs have broad applications across industries, including:

• Automated customer support in multiple Indian languages.

• Voice recognition systems for accessibility and inclusion.

• Language translation tools for businesses and government agencies.

• AI-powered legal drafting and data extraction, through the A1 generative AI tool.

Additionally, the affordability of AI voice agents (priced at Rs 1 per minute) makes them accessible for sectors like healthcare and banking, where multilingual support is crucial.

Challenges and Future Trajectory

Struggles in gaining industry traction

Despite raising $54 million for AI development, Sarvam AI faces challenges in establishing itself within India’s AI ecosystem. Some of the key hurdles include:

• Limited adoption within India’s AI community, despite technological breakthroughs.

• High computational costs, requiring extensive infrastructure investments.

• Competition from global AI giants, such as Google’s Gemini and OpenAI’s GPT models.

Future roadmap

Sarvam AI aims to expand its AI capabilities by:

• Enhancing Shuka’s capabilities with future versions of Llama.

• Adding support for more Indic languages, increasing inclusivity.

• Scaling up datasets and training resources, improving model accuracy.

• Strengthening partnerships with AI research organizations like AI4Bharat and Nvidia.

As India continues to embrace AI-driven solutions, Sarvam AI’s innovations have the potential to shape the future of multilingual voice AI and sovereign generative AI for Indian users.

Conclusion

Sarvam AI’s development of Shuka v1 and Sarvam 1 marks a significant leap in India’s AI landscape. By creating open-source, voice-first AI models, the company is addressing a crucial need for regional language accessibility in AI applications. Despite challenges, Sarvam AI’s mission to develop full-stack generative AI solutions tailored for India places it at the forefront of multilingual AI innovation.

As newer versions of Llama and expanded datasets become available, Sarvam AI is well-positioned to advance AI-driven communication in India, ensuring that AI technology is inclusive, accessible, and secure for millions of users.

The gold standard of business news

Morning Brew is transforming the way working professionals consume business news.

They skip the jargon and lengthy stories, and instead serve up the news impacting your life and career with a hint of wit and humor. This way, you’ll actually enjoy reading the news—and the information sticks.

Best part? Morning Brew’s newsletter is completely free. Sign up in just 10 seconds and if you realize that you prefer long, dense, and boring business news—you can always go back to it.

Join 4.3 Million Readers Now

Flipped.ai: Revolutionizing Recruitment with AI

At Flipped.ai, we’re transforming the hiring process with our turbocharged AI recruiter, making recruitment faster and smarter. With features like lightning-fast job matches, instant content creation, CV analysis, and smart recommendations, we streamline the entire hiring journey for both employers and candidates.

For Companies:
Looking to hire top talent efficiently? Flipped.ai helps you connect with the best candidates in record time. From creating job descriptions to making quick matches, our AI-powered solutions make recruitment a breeze.

Sign up now to get started: Company Sign Up

For Job Seekers:
Explore professional opportunities with Flipped.ai! Check out our active job openings and apply directly to find your next career move with ease. Sign up today to take the next step in your journey.

Sign up and apply now: Job Seeker Sign Up

For more information, reach out to us at [email protected].

Want to get your product in front of 75,000+ professionals, entrepreneurs decision makers and investors around the world ? 🚀

If you are interesting in sponsoring, contact us on [email protected].

Thank you for being part of our community, and we look forward to continuing this journey of growth and innovation together!

Best regards,

Flipped.ai Editorial Team