Whisper's output for Hindi audio leads to a transcript in Arabic or Persian script

[abhilashgupta]

I tried transcribing a Hindi (non standard dialect) audio file.
The API call went through fine and returned the transcription in the Devnagiri script.
However, when I run whisper transcribe on it locally, it recognises that the audio is in Hindi, but the transcription is in either Arabic or Persian script (or maybe Urdu, not sure).
Same when I force it to transcribe in Hindi, Nepali or Marathi language (almost same script as Hindi, one extra character).

Writing here because I couldn't see an option to raise an issue.

Steps to reproduce:
I cannot attach a .wav file here, so here is the link to the video -> https://www.youtube.com/watch?v=x2CnZZFf8Jg
To use API, call the API and get back response in correct script.
To run locally, just run whisper downloaded_audio.wav --language hi --model base or whisper downloaded_audio.wav --model base

There's a conversation about how Whisper's auto detection detects Hindi audio to be Urdu (which makes sense given their similarity) at #118 but the API gives the response in the correct script and the toolkit outputs it in the wrong script even if I explicitly mention the language

[7gxycn08]

Don't bother using Whisper the accuracy rate is pretty low compared to letting GPT3.5 do the translation.

[abhilashgupta]

I am not translating here.
I have an audio file from which I want to extract the transcription of the audio in the same language, which is a very different problem than translation, no?

[kennethleungty]

@abhilashgupta Did you come across any solution/new insights for this?

[abhilashgupta]

Yes,
the issue was observed in the base model, however it isn't the case with the large (large-v2 or large-v3) models. So, I started using the large models wherever possible.
However, occasionally, it still detects the language Hindi as Urdu, when set to auto-detect, given that they both are close to each other, and in that case, it still returns the Urdu script transcriptions. For that, always set the language manually to Hindi where possible.

[Aliraza1010a]

Bhai me bhi ye issue face kar raha hu
Please batao kaise isse fix kare

[nairajay2k]

I am facing the same issue. When I call model.transcribe('filename', language='hi') the returned text is in urdu/persian.

[gaurav241102]

did you come up with any solution for this?

[firofame]

same issue with Malayalam audio

model = whisper.load_model('large-v3')
result = model.transcribe("audio.mp3")
print(result["language"]) // correct language detected i.e "ml"
print(result["text"]) // text is in a different language