Real-time Speech to Text Converter and Communication Coach for Low-Resource Language (Hindi): Sanchar Shuddhi

This paper presents the development of a real-time application that converts Hindi speech to text and provides in-conversation suggestions to call center agents, addressing communication challenges in low-resource languages. The system comprises of five core modules: Speaker Diarization, Automatic Speech Recognition (ASR), Sentiment Analysis, Grammar Correction, and Sentence Rephrasing. By leveraging Apache Kafka, the application ensures efficient load balancing and modularity by creating distinct topics for each module. During calls, agents receive real-time guidance, while post-call analysis refines transcripts by correcting grammar and rephrasing sentences, creating a valuable reference for future interactions. This dual approach enhances both immediate and long-term agent performance. The proposed solution highlights the potential to improve customer satisfaction and agent efficacy in Hindi call centers. It lays the groundwork for future enhancements, such as multi-language support, improved sentiment analysis, and integration with Customer Relationship Management (CRM) systems for streamlined workflows.The models used to achieve ASR was Whisper Large-v2 and Llama 3.1:8b for the grammar correction and rephrasing. The ASR achieves a correctness rate of 73.6%, successfully identifying 1233 out of 1784 words in the test set.Overall Diarization error rate is 33.91 % achieved. Sentiment analysis module achieved overall accuracy of 71%.