What Can Local AI Do Besides Text? Vision, Speech and Images
A local setup can read images, transcribe audio, speak text and draw pictures. Each skill uses a free model and a small free tool on your own machine.
The main idea
Most people use local models only to chat. In the video I add four more abilities to one machine, all free and all private.
First, vision. I pull Gemma 4 E2B in Ollama, upload a picture and ask what the person is wearing. It answers fast, even though it cannot say who the person is. Second, hearing. Whisper is OpenAI's open-source transcription model. It handles 99 languages. After installing ffmpeg and Whisper, one command with the turbo model transcribes a file with timestamps.
Third, speech. Kokoro is an 82 million parameter model that runs on a plain CPU. The model files together are under half a gigabyte. I feed it a text file and get a WAV file back. Fourth, images. On a Mac I use Diffusion Bee, and I type a prompt to get a picture. On Windows the free equivalent is Fooocus. Every one of these is a model file plus a small free tool on hardware I already own.
Key points
- Gemma 4 E2B in Ollama answers questions about an uploaded image on your own machine.
- Whisper handles 99 languages; on a Mac, brew install ffmpeg comes first, then one command with the turbo model transcribes a file.
- Kokoro has 82 million parameters, runs on a CPU and needs under half a gigabyte of model files.
- Diffusion Bee on Mac and Fooocus on Windows generate images from a text prompt for free.
- Hugging Face lets you browse models by tag, such as TTS, and each model card lists the license, size and how to run it.