Please we want an Arabic Specialized speech to text model.

[alaya-source]

We want a model that specializes for speech to text on Arabic language only the current model does not provide an accurate and good Arabic experience.

[AlexanderYastrebov]

Here is an overview of the current state Open Universal Arabic ASR Leaderboard

If I read it correctly the best Word Error Rate is ~30% (compared to ~6% for English)

Links:

• https://github.com/Natural-Language-Processing-Elm/open_universal_arabic_asr_leaderboard

@BenoitWang Hello, thank you for creating Open Universal Arabic ASR Leaderboard. Did I get it right that current SOTA for Arabic has way higher error rate than for English and European languages? Could you maybe suggest a model suitable for inclusion into Handy? Thanks.

[BenoitWang]

Hi @AlexanderYastrebov , the hard part of Arabic ASR is that it's so rich in dialects and they do vary from each other. 30% WER is about one universal model dealing with all dialects. Can consider any Top models on the leaderboard.

[AlexanderYastrebov]

Previous #495

[wolfenazz]

That’s why it’s considered one of the most challenging languages in the world. Still, we need a speech-to-text model that proves otherwise! ^_^

[cjpais]

The latest cohere model supports, I just merged support in #1200

I do not know how good it is for Arabic, but maybe give it a try. Right now this model is running on the CPU only so it might be a bit slow, in the future especially for Mac's we will be adding GPU acceleration but for now we at least support it.