New AI Translates Arabic Dialects Better Than Ever

A research group in Abu Dhabi called the Technology Innovation Institute (TII) has created a new AI tool called Falcon-ASR. It listens to speech and writes it down as text. The special thing is that it's really good at Arabic, especially the Emirati dialect that people actually speak every day. Most AI struggles with dialects because they're not written down as much.

Falcon-ASR was trained on lots of different kinds of Arabic, plus English, French, Spanish, and Portuguese. It can handle background noise, people talking over each other, and phone-line audio. In tests, it made fewer mistakes than other top systems. For example, when transcribing Emirati speech, its error rate was about 23%, which is better than the next best system by a noticeable amount.

The model also gives you timestamps for each word, so you know when it was said. This could be useful for creating subtitles or searching audio. The team has a demo you can try online, and they plan to offer API access and apps later. So if you work with Arabic audio, this could save you hours of manual transcription. But it's not perfect: it still makes mistakes, especially in noisy conditions or with very rare dialects. And it's not yet available as a finished product, just a demo.