Local audio models

Each page below covers one model family you can download and run on your own hardware, with no account and no upload of your recordings to anyone else.

What a model page here answers

Every model page answers the same four questions in the same order, so you can compare two of them without reading either in full.

What it is and who makes it. The organisation behind the weights, what the model was trained to do, and the licence shape in broad terms.
Sizes and hardware. Which variants exist, roughly how heavy they are, and whether you need a GPU or can get away with a CPU. Hardware pages are linked where a specific card matters.
How people run it. The general shape of the tooling: what kind of program loads the model, what it expects as input, what it hands back.
Fit. What the model is genuinely good at, and the jobs where it is the wrong tool and you should reach for something else.

No benchmark tables and no speed claims. Audio quality is subjective and hardware varies too much for numbers to travel well, so these pages describe shape and tradeoffs and leave the measuring to you on your own machine.

Speech to text

Whisper. Transcription and subtitling, in a ladder of sizes from tiny to large. Handles many languages, and runs acceptably on CPU at the smaller sizes.

Text to speech

Piper. Small, fast, offline speech synthesis with a large library of per-language voices. Built to run on modest hardware including single-board computers, which is why it turns up in local assistants and screen readers.

Music and sound generation

MusicGen. Text-to-music generation that produces short instrumental clips from a written description. The heaviest of the three by some distance and the one that most wants a GPU.

Choosing between them

These three do not compete with each other. They do different jobs, and a full local audio setup often runs two or three of them side by side: Whisper to turn recordings into text, Piper to read text back out, MusicGen when you need a bed of music that nobody owns.

So choose by job rather than by model. Work out whether you want to understand audio, produce speech, or invent music, then take the page above that matches. The getting started guide walks through that decision and what to expect from a first run.