DeepSDR Updates: Live Public Safety Incident Map with an RTL-SDR, Whisper and an LLM

Last week we posted about Hubquhq's DeepSDR project (project contact page), which collects transcribed voice data from multiple public safety radio channels and then uses an LLM to summarize and plot incidents on a map.

Recently, he wrote in again and wanted to share additional information about the project and how it works after receiving a few questions from interested parties. He writes:

I run every stage in a separate docker container: Capture (I capture audio as longer sessions for archiving, and separately just short sessions where there's no silence for further processing. I tried using VAD to detect voice, but with radio it was missing some talk, so I just use ffmpeg) > Transcribe > LLM > Finding latitude and longitude for addresses > Grouping connected conversations. 

He has also provided a PDF explaining how DeepSDR works in more detail, which you can download here.