Imagine a world where language learning apps could truly understand and assess not just how you pronounce words, but also how well you use grammar when you speak. This isn't science fiction; it's the focus of exciting new research that's pushing the boundaries of spoken language assessment. Traditionally, evaluating spoken language skills has focused on pronunciation and fluency, leaving grammar to written tests. This separation creates a gap between how we communicate in real life and how we're tested. This new research introduces an innovative system that leverages the power of large language models (LLMs) to assess grammar directly from spoken language. Here's how it works: an LLM generates unique paragraphs with specific grammar points to be evaluated. The student reads the paragraph aloud, and their speech is analyzed using a custom-built speech recognition system, which is key to accurately capturing the nuances of spoken grammar. The system then compares the student's speech to the grammatically correct version, providing a score that reflects their mastery of the targeted grammar points. What makes this approach revolutionary is its ability to create a virtually endless supply of unique test paragraphs. This prevents students from simply memorizing answers, promoting genuine learning and making the assessment more robust. The research also highlights the importance of specialized language models within speech recognition systems. Standard speech recognition often struggles with grammatically incorrect sentences, as it's trained on primarily correct text. This new system, however, incorporates a custom language model trained on a wider range of grammatical variations, leading to significantly improved accuracy in capturing spoken grammar. The results are impressive: the custom system far outperforms standard speech recognition in correctly transcribing and evaluating grammar in spoken responses. This innovative approach opens doors to a more comprehensive and personalized language learning experience, offering real-time feedback on spoken grammar and paving the way for more effective language learning tools in the future. While challenges remain in terms of dataset availability for training and evaluation, this research marks a crucial step toward AI-powered grammar assessment that truly reflects real-world communication.