✈️ Telegram developer dkaraush has released tinyasr—a speech recognition model that runs directly on the device. You can try it out in your browser; the audio isn’t sent anywhere.

The model is ~32 MB in size and contains 63 million weights. It makes mistakes on 5.5% of words in Russian and 6.1% in Ukrainian. It also supports English, Spanish, Italian, Portuguese, Uzbek, Persian, and Khmer.

#Telegram