Improve Language detection

[ab-pandey]

Since wishper detects language based on first 30 secs of the audio, sometimes there are errors in language detection. For example
This video is in english but both whisper and faster_whisper detects language as hindi "hi". There is a solution for this issue in whisper here. Since I am new to Ctranslate2, I am having difficulty in cloning this solution to faster_whisper. Can someone help me on this? Thanks.
P.S. I have tested the solution in whisper and it works.

[guillaumekln]

Since I am new to Ctranslate2, I am having difficulty in cloning openai/whisper#676 to faster_whisper. Can someone help me on this? Thanks

What have you tried so far? Maybe you can share your current changes and we can help you implement this.

[ab-pandey]

Actually I am not able to figure out how to proceed further. There are several variables in the fix like "content_frames", "N_FRAMES", "segment"; functions used "pad_or_trim" in this solution which I am not able to figure out how to get the correct values for faster_whisper implementation

[elloza]

has been this implemented?

[dilerbatu]

Any update ?

[trungkienbkhn]

FYI, I created a new PR to implement this feature: #732

[ldolegowski92]

Do Twojej wiadomości, utworzyłem nowy PR, aby wdrożyć tę funkcję: #732

How will the model behave if the conditions are not met?