An AI Driven Voice Enabled System for Automated Resume Aware Interview Question Generation and Semantic Response Evaluation Based on Comprehensive Parameters for Large Language Models

The increasing competitiveness of modern recruitment and assessment processes has created a growing demand for intelligent systems that support effective interview preparation and automated response evaluation. Traditional platforms rely on static question repositories and rule based evaluation, limiting personalization, adaptability, realistic interaction, and consistent feedback. Traditional platforms further depend on structural parameters such as job roles, descriptions, skills, candidate responses, context, and behavioral metrics, lacking semantic understanding and effective resume integration. To overcome this limitation, VELIS (Voice Enabled LLM based Intelligent Interview System), an artificial intelligence driven and resume aware system, automatically generates interview questions and evaluates candidate responses through advanced natural language processing techniques and Large Language Models. It advances beyond existing systems by introducing a comprehensive set of parameters including authentication details, resume, spoken responses, resume features, experience, context, and semantic similarity, enabling adaptive, personalized, and semantically consistent generation of interview questions, assessments, and feedback. The process is modeled through the integration of multiple components within VELIS, where the OpenAI API with GPT-4 generates context-aware interview questions and performs response evaluation, BERT extracts structured resume information including skills, technologies, projects, and experience, Whisper enables speech-to-text transcription, Tacotron-2 generates natural speech output for real-time interaction, and Sentence-BERT evaluates semantic relevance. Experimental evaluation shows 0.94 question relevance accuracy, 0.91 speech recognition accuracy, 0.08 transcription error rate, 0.88 speech naturalness, 0.91 response evaluation F1 score, 0.93 overall accuracy, and 0.82 seconds latency, demonstrating efficient automated interaction and improved interview preparation.Code and data are available at github https://github.com/poornarc/VELIS/tree/v1.0 and DOI https://doi.org/10.5281/zenodo.19332758.