[Q] Silence audio gives random text

[henrysydney]

I run some tests with latest versio stable-ts today, but I face this issue with most silence audio and it returns to random texts. I tested on same model tiny-whiser english on HuggingFace and it returned a very good result. Not sure, if the model on whisper github vs. huggfing face are different. I attached audio and screenshot below.

Audio link: sample_audio.wav

• stable-ts with lastest whisper-tiny english model, it results text: " Galglic! hello welcome helloகly Mortal Kombat Final Fantasy"

• I tested again stable-ts with lastest whisper-tiny english model, it results different text: " See how they lower.",

• I tested this audio also on huggingface tiny-english whisper: "hello hello hello"

I'm newbie on this, I'm working around to fix this issue before implementing to my project for help my student to practice English in class.

Thanks so much for your contribution to the community @jianfch

[jianfch]

Unless the huggingface model you used has been finetuned or retrained, it might be just due to the differences in the default arguments. Using temperature=0.0 with stable-ts on the tiny English model, I got: " Hello, hello, hello.".

[chislon]

What parameters are you using with the file?