AI-Powered-Meeting-Assistant

The Problem

Manual meeting note-taking consumed significant time while missing key action items and decisions requiring follow-up.

The Solution

Built a modern meeting intelligence platform that transcribes live audio and generates actionable summaries using state-of-the-art NLP models. Implemented real-time speech-to-text processing with Python and Flask, reducing manual note-taking time.

Impact

Intelligent meeting transcription and summarization reducing manual documentation overhead.

Architecture

Key Challenges

• Whisper transcription accuracy dropped significantly with overlapping speakers and background noise. Implemented a pre-processing step using WebAudio API to apply noise suppression before sending audio chunks, which improved transcription accuracy on noisy calls noticeably.
• Processing 30-second audio chunks created visible gaps in the live transcript display. Solved by overlapping chunks by 3 seconds and deduplicating the overlap region using word-level timestamp comparison from Whisper's verbose output.
• Action item extraction was unreliable when meeting topics shifted mid-conversation. Switched from a single end-of-meeting prompt to a rolling summarization approach — summarizing every 10-minute window and merging summaries — which preserved context across topic changes.

Key Learnings

Audio quality is the ceiling for any speech-to-text pipeline — no amount of prompt engineering compensates for poor input audio. I also found that action item extraction is a harder NLP task than summary generation: humans rely heavily on tone and implicit commitment signals that LLMs miss without explicit instructions to look for them.