Towards Reliable AI-Based Assessment: Modular Scoring and Hallucination Control

With the rapid expansion of digital education, there is a growing need for scalable and objective methods to evaluate subjective responses, particularly spoken, responses-areas where traditional manual grading often proves inconsistent and resource-intensive. This research introduces a voice-enabled, AI supported evaluation framework specifically designed to automate the assessment of oral responses. The proposed system integrates speech recognition, context-aware answer generation using Retrieval-Augmented Generation (RAG), and a multi-stage evaluation pipeline to ensure reliable grading. The architecture employs Sentence-BERT embeddings for contextual retrieval, FAISS indexing for efficient similarity search, and transformerbased models to generate responses grounded strictly in the retrieved context. The evaluation process is structured across four distinct stages to effectively reduce hallucination and bias. It begins by identifying the question's core concepts, then measures semantic similarity using embeddings, analyzes logical coherence, and finally assesses language quality. This multi-layered approach feeds into a dynamic scoring system that adapts to question difficulty and answer depth. The system's real-world utility is showcased through a Streamlit interface that supports voice input and delivers immediate feedback. Experimental results across varied academic domains show improved grading consistency, contextual relevance, and overall transparency compared to standard LLM-based or rule-based assessment approaches. This modular, scalable framework provides a robust foundation for integration into learning management systems, remote education tools, and recruitment platforms, advancing a more ethical and efficient standard for automated oral assessment.