Jakub Cerovský

Side project · Ongoing

Local AI lab on Apple Silicon

Running voice cloning, lip-sync video, and local LLMs entirely on a MacBook Pro M4 Pro - no cloud required.

Stack

Apple SiliconLocal LLMs (GGUF/MLX)LivePortraitFaceFusionVoice cloningPython

Overview

An ongoing side project: how much of the modern AI media stack can run entirely on a 24 GB MacBook Pro M4 Pro? The answer is a surprising amount, with careful tool selection.

On the voice side - Slovak voice cloning, a niche most tools don't support. That meant testing recording pipelines, building a phonetically rich Slovak training script, and setting up real-time voice conversion on macOS with virtual audio routing.

On the video side - evaluating the field of face animation and lip-sync tools (SadTalker, Wav2Lip, MuseTalk, LivePortrait, LatentSync and others) to find which actually run on Apple Silicon rather than assuming a CUDA GPU. And on the text side - benchmarking quantised local LLMs against the machine's real memory bandwidth to find what is genuinely usable, not just what technically loads.

Outcome

  • A working local pipeline for Slovak voice cloning and talking-head video.
  • A practical shortlist of AI tools that genuinely run on Apple Silicon.
  • Hands-on understanding of quantisation, memory bandwidth and MoE models that transfers directly to professional AI work.

Related