Authors - Milind Nemade, Khush Chheda, Rahul Dhanak, Durgeshkumar Dubey Abstract - The problem of meeting productivity continues to prevail in the current era in multilingual environments due to frequent language switching between speakers. Most of the existing frameworks for meeting intelligence primarily focus on automatic transcription and lack significant support for Indic languages, speaker identification, and task extraction. Additionally, many of these frameworks depend on metadata associated with specific platforms and, therefore, cannot be used in any offline environment or even on other platforms. In this work, we present a scalable and platform-agnostic framework for meeting intelligence which can automatically an alyze meetings post factors by leveraging speech recognition, speaker identification, and contex tual analysis. The system leverages multilingual and code-switched transcription capabilities of Sarvam AI, generates speaker embeddings using ECAPA-TDNN, and then uses Large Language Models for context-based analysis. Two different strategies for speaker identification are dis cussed in this paper such that they do not need any platform-based metadata while improving the attribution accuracy. We have further developed an asynchronous framework for extracting tasks, assigning tasks, and notifying about them. Experimental results indicate enhanced transcription accuracy as well as speaker identification accuracy in Hindi-English code-switching cases. Fu ture work will focus on implementing advanced privacy protection and end-to-end encryption mechanisms for secure storage and processing of meeting recordings and metadata.
Open Zoom