BharatGenAI introduced the multilingual speech recognition model Shrutam-2

Developers from BharatGenAI released the second version of the automatic speech recognition (ASR) model called Shrutam-2. The system is designed to process audio streams in 12 Indian languages, providing high transcription accuracy in a multilingual environment. The model's architecture is based on LLM principles and uses a specialized Mixture-of-Experts (MoE) projector. This approach allows for efficient allocation of computational resources when processing various linguistic structures characteristic of India's regional languages. The model demonstrates a result of 24 points on the internal quality metric with a processing time of 2 hours on the test dataset. The project is available for download on the Hugging Face platform. The use of Shrutam-2 is limited by a non-commercial license, implying the application of the technology exclusively for research or educational purposes. The developers provided access to the model's weights for integration into local speech processing pipelines. The release of this model expands the toolkit for working with low-resource languages, where standard global ASR solutions often show insufficient effectiveness. The transition to the MoE architecture in speech recognition tasks allows optimizing system performance without the need for a multiple increase in the number of parameters, which is critical for deployment on specialized equipment.