AI arm of Sony Research to help develop large language model with AI Singapore

Sony Research has inked a partnership to help test and finetune the Southeast Asian Languages in One Network (SEA-LION) artificial intelligence (AI) model, focusing on Indian languages. 

The AI arm of Sony Research will work with AI Singapore (AISG) responsible for the development of SEA-LION, to plug gaps in ensuring the large language model (LLM) stands up well on the global landscape, representing the region’s populations and languages. The partners said in a statement Tuesday their research collaboration will involve LLMs under the SEA-LION umbrella, all of which are pre-trained and instruct-tuned specifically on Southeast Asian cultures and languages. 

The open-source LLM has been trained on 981 billion language tokens, which AISG defines as fragments of words created from breaking down text during the tokenization process. These fragments include 623 billion English tokens, 128 billion Southeast Asia tokens, and 91 billion Chinese tokens.  

Also: Lost in translation: AI chatbots still too English-language centric, Stanford study finds

The partnership entails that Sony will work on tests and feedback on the AI model, tapping the Japanese vendor’s research presence in India and expertise in the development of LLMs for Indian languages (including Tamil). Tamil is estimated to be used by 60 million to 85 million people globally, most of whom are based in India and Southeast Asia. 

Sony will exchange best practices on LLM development and research methodologies, as well as the application of its research in speech generation, content analysis, and recognition. 

The integration of the SEA-LION AI model with Tamil language capabilities has the potential to boost the performance of new applications, said AISG’s senior director of AI products Leslie Teo. He added that the Singapore agency will also share its knowledge and best practices in LLM development. 

Also: AI leaders urged to integrate local data models for diversity’s sake

IBM and Google are among other industry players drawn into finetuning the regional LLM, including making it available for developers to build customized AI applications. 

“Access to LLMs that address the global landscape of language and culture has been a barrier to driving research and developing new technologies that are representative and equitable for the global populations we serve,” said Hiroaki Kitano, president of Sony Research. “Diversity and localization are vital forces. In Southeast Asia specifically, there are more than 1,000 different languages spoken by the citizens of the region. This linguistic diversity underscores the importance of ensuring AI models and tools are designed to support the needs of all populations around the world.”

Also: Transparency is sorely lacking amid growing AI interest

Established in April 2023, Sony Research focuses on technological development that can improve content creation and fan engagement, including in the areas of AI, sensing, and virtual spaces. For instance, its deep learning research team has been working on technologies that encompass, among others, model compression and neural rendering, which it hopes can be integrated into Sony’s GUI development tool Neural Network Console, and open-source libraries Neural Network Libraries. 

These technologies can be used in AI-powered electronics products spanning various sectors, such as games, movies, and music, and games, Sony said. 

Its interactive entertainment unit has filed a patent for a “harassment detection apparatus” that includes an input unit built to receive biometric data and with capabilities to generate, based on biometric data, emotion data associated with users, according to an April 2024 publication on World Intellectual Property Organization’s PatentScope search platform.

Also: New global standard aims to build security around large language models

With the system, Sony hopes to be able to detect and mitigate communications between individuals in multi-player games or virtual reality experiences that are malicious, such as harassment. Tapping machine learning and AI models, the system can detect biometric data such as speech and determine a player’s emotional state, for instance, through sounds such as sobbing and screaming. These may be used to identify victims of harassment within the shared environment, according to the filing. 

In May, Sony Music Group released a statement noting that its artists’ copyrighted works, including compositions, lyrics, and audio recordings, should not be scraped and used to train AI models unless explicitly authorized.