همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Speech-Based Conversational Business Intelligence System Leveraging Large Language Models for Accounting Analytics

نویسندگان :

Fatemeh Mohsennia¹ Mohsen Kahani² Morteza Fardin³

1- دانشگاه فردوسی مشهد 2- دانشگاه فردوسی مشهد 3- شرکت فناوری اطلاعات ژرف پویان باران

کلمات کلیدی :

Conversational Business Intelligence،Automatic Speech Recognition،Natural Language Processing،Large Language Models،Accounting Analytics

چکیده :

Abstract— In this study, a speech-based Conversational Business Intelligence (C-BI) system was designed and implemented to facilitate natural interaction between users and accounting software. The proposed system provides a seamless communication channel that enables non-technical users to obtain financial reports and analytical insights through Persian speech or natural language queries, without navigating complex interfaces. The architecture integrates two advanced components: an Automatic Speech Recognition (ASR) model based on Whisper, fine-tuned with domain-specific Persian accounting data, and a Large Language Model (LLM) responsible for rewriting, interpreting, and extracting key concepts from user queries. In addition, the BGE-M3 embedding model was employed to generate semantic representations of queries and perform similarity-based retrieval within a vector database. Experimental results demonstrated that the fine-tuned Whisper model (v2) achieved a Word Error Rate (WER) of 0.11 and a BLEU score of 0.817, while the Gemma3--12B LLM reached an overall accuracy of approximately 90% in intent classification and entity extraction. These findings confirm that combining speech and language models within the Conversational BI framework provides an effective and localized solution for improving user experience in Persian accounting systems and enhancing the accessibility of financial data analytics.