whisper-jax icon indicating copy to clipboard operation
whisper-jax copied to clipboard

Whisper JAX incorrectly detects the language

Open realfolkcode opened this issue 1 year ago • 0 comments

I want to transcribe a musical educational video (e.g., this). To do this, I am using a JAX gradio endpoint API. But it fails to correctly detect the language (which is English, in my example) and instead it outputs a text that seems to be in Slovenian. I am familiar that this problem has already been acknowledged previously. However, I dug a little deeper and discovered that Whisper from transformers is less prone to mixing languages. This might indicate towards potential bugs in JAX implementation.

Here is a colab notebook that showcases the differences between JAX and transformers outputs.

realfolkcode avatar Sep 22 '23 08:09 realfolkcode