> recording, transcribing and taking minutes of all meetings with AI. Then that gets fed into a RAG system and we can see what's been said when.
Could you please share details on tooling for this?
Could you please share details on tooling for this?
I use audacity to capture the audio stream live, whisper.cpp with person segmentation to extract the text with notations on who was talking and a fine tuned llama model which is trained on data specific to the companies market to extract the minutes. It's a couple of passes for the model to write the prompts to extract the data into minutes.
The RAG system is just untrained sbert weights of every sentence in the recording and minutes shoved into the vector postgres. Queries are @mentions in the company chat.
Basically a bunch of python scripts held together by duct tape and chewing gum.
That's got nothing to do with the LLM used to do the summary but part of the RAG system.
To be clear, my issue was I could only capture the streaming audio of my mic or the people on the call. Never both at the same time.
I just have a dummy user that joins the meetings so people known when they are being recorded. I imagine I accidentally dodged a bullet there since that user isn't supposed to have audio in and the audio out is directed to audacity and muted otherwise.
> whisper.cpp with person segmentation
Super curious about this bit. Did you use the --diarize or the --tinydiarize flag? How accurate/reliable was the segmentation? Sounds like it was good enough to use at the very least. I had looked into this a few months ago but thought it wasn't good enough yet.
Jokes aside, excel is the most powerful business software ever. Everyone use it.
Maybe it works better with an American accent.