Real-Time Call Analysis Dashboard

An automated pipeline for transcribing calls and performing sentiment analysis. Azure Speech streams real-time transcriptions, then an LLM analyzes the call and generates a report on agent performance.

Tech Stack

• Python
• Azure Speech Services
• Azure AI Foundry (DeepSeek V3-2)
• Streamlit

How It Works

• Input: Upload a .wav file or stream live audio
• Transcription: Azure Speech streams real-time partial and final results
• Analysis: DeepSeek V3-2 performs call breakdown and identifies coaching opportunities
• Output: JSON file with full transcript and structured feedback

I used a detailed system prompt to control the LLM output, plus the max_tokens parameter to adjust report length. Caching prevents redundant API calls when the same file gets uploaded twice.

Demos

Terminal demo:

Streamlit interface with real-time transcription streaming:

What I Learned

Pairing specialized Azure services with LLMs works well. Azure Speech handles transcription, the LLM adds analysis. Together they do more than either could alone.

Real-time streaming makes a big difference in user experience. Watching transcription appear word-by-word feels more responsive than waiting for batch processing. The caching was a practical addition that saved API costs on duplicate uploads.