CS Master's Thesis

2026-03-13

Research / thesis

Master's thesis: design and formal evaluation of a local, modular voice assistant with self-hosted language models, local speech services, and episodic memory.

research · thesis · llm · voice · latex

Repository

This repository holds the LaTeX source for my master’s thesis. The thesis and AI Assistant are one project from two angles: the thesis argues and measures; the codebase runs.

What the thesis formally contributes

  • A reference architecture for a local, modular voice assistant with stable contracts between STT, LLM, and TTS.
  • A benchmark campaign covering 78 LLM data points, 14 STT data points, and 24 retrieval-overhead measurements across x86 and ARM platforms.
  • A result that memory mode shapes end-to-end latency more than model swapping within a weight class — with an extreme case of 2.4x median latency increase when retrieval is enabled.
  • An honest scope boundary: the evaluation covers operational latency and architecture trade-offs, not semantic quality of LLM outputs or perceptual TTS quality.

Where to read more

The full technical account — architecture, deployment, pipeline, memory design, and what the benchmark numbers mean — is in Local voice assistant: architecture, deployment, and what the benchmarks found.

The implementation side is in AI Assistant.

The citable PDF and full tables are in the cs-masters repository.