Free AI Tool Now Transcribes Your Audio Almost Perfectly

A popular free AI tool called llama.cpp just released a small but meaningful update. llama.cpp (a free program that runs AI models directly on your own laptop or phone) fixed a bug in how it handles audio for one specific voice model, called LFM2. The bug was in the "preprocessor" — the part that cleans up sound before the AI tries to understand it, like a hearing aid tuning the audio before your brain interprets words.

The bug was not tiny. On test recordings, the tool got about 4.5% of English clips wrong and 6.5% of Japanese clips wrong. In Japanese, some mistakes changed entire words — the kind of error that could turn one meaning into a completely different one. After the fix, accuracy jumped to essentially perfect: 200 out of 200 English clips correct and 200 out of 200 Japanese clips correct in the tests the developers ran.

Why should you care? Because this is free, private voice-to-text. Unlike cloud services that send your recordings to a company's servers, llama.cpp runs entirely on your own device. That means your meeting notes, voice memos, or interviews never leave your computer. Nothing to subscribe to, nothing uploaded. It works on Windows, Mac, Linux, Android, and even Apple Silicon and Snapdragon chips. The same update ships across all of them.

The catch: this fix only improves LFM2, one specific audio model — other audio models in the tool are unchanged. And you still need a reasonably modern computer and a bit of setup patience to run it yourself. It is not a one-click app for everyone yet. But for anyone who wants accurate, free, private transcription, this is a genuine step forward.