Speech Analytics on a Home PC

The author shares their experience building a local pipeline for analyzing phone calls using a single gaming graphics card. The article covers the technical implementation of a system capable of processing audio streams without sending data to cloud services. Special attention is paid to the use of LLMs in tasks where classical speech recognition—identifying 'who said what'—is secondary to extracting semantic entities and analyzing dialogue context. The author details the solution's architecture, tool selection, and the optimization of computing resources for home hardware. This article is useful for developers interested in MLOps, local audio processing, and implementing neural network solutions for communication analysis tasks.