GigaAM: Speech to Text With No Internet

Most people talk faster than they type. Saying a thought takes seconds, typing it takes a minute. We’re just used to text being something you type, so the gap never registers.

That’s where most of my writing starts now: notes, messages, half my email.

It all used to go through the cloud. Record, upload, wait, get text back. Fine, as long as you have a connection and you don’t mind your voice sitting on somebody else’s server.

Now my offline mode runs on GigaAM v3, a speech recognition model Sber released openly.

I hold a hotkey, talk, let go, and the text is on screen. Nothing gets uploaded and there’s nothing to wait for. The model sits on disk and the audio goes straight into it.

The Russian is what surprised me. Most models are trained mostly on English and Russian comes along for the ride, which you hear on every name and every technical term. GigaAM was built for Russian first and you can tell within a sentence.

It usually punctuates on its own too, so what comes out is finished text rather than a wall of words.

But the part I got used to isn’t the speed. It’s not having to think about what I’m dictating. Client details, half-formed ideas, somebody’s phone number, all of it stays on the machine. It also works on a plane and on bad hotel wifi, doesn’t meter minutes, and won’t break when someone changes their pricing.

The downsides are small. The model is 215 MB, an int8 build, so you never notice it on disk. Other models do better on English, which is what I kept the cloud for.

All of it lives in CrookedVox: hold a key, talk, get text, push it through AI if you want it fixed up, shortened or rewritten in another tone. The cloud path is still in there, I just reach for it less and less.

I’ve been using this every day for a few weeks. I don’t miss the cloud.