Issue with Arabic Transcription in Whisper-Large-V3-Turbo ("نعم" → "Naah"/"Naahe")

Community
, ,
1

I’m using whisper-large-v3-turbo to transcribe voice inputs in both English and Arabic. However, I’m encountering an issue where the Arabic word “نعم” (which means “yes”) is consistently being transcribed incorrectly as “Naah” or “Naahe”.

Has anyone else experienced this behavior with Whisper? If so, what strategies or configurations have you found effective in improving transcription accuracy for short Arabic words like this?

Any insights or suggestions would be greatly appreciated.

1 Like
2

You can try a different model.

While whisper has less truncation problems, other models can perform better for specific languages.

https://openai.com/index/introducing-our-next-generation-audio-models/

This bar chart compares the word error rates of four speech-to-text models across various languages, showing that newer models generally achieve lower error rates, especially in widely spoken languages. (Captioned by AI)
This bar chart compares the word error rates of four speech-to-text models across various languages, showing that newer models generally achieve lower error rates, especially in widely spoken languages. (Captioned by AI)1508×843 47.2 KB

Also, all models perform significantly worse than usual if the audio is too short with only a single word.

1 Like
3

My priority is to use a free model for transcription. Can you suggest any?

4

Nothing that comes to mind ATM, but I’ve heard some people do fine-tuning on whisper to improve performance.

1 Like
5

Thanks for flagging this. Can you please generate a HAR file as you experience this error and email those details to support@openai.com?

6

I had the same issues, but when I played with the parameters of my voice activity detection algorithm and the transcribed chunks sizes, it got fixed. One more important thing is the temperature, when you reduce it to around 0.1, this issue happens less often as I experienced.

Good Luck!