different behavior dealing with Japanese Kanji by different model #204
-
When transcribing some Japanese language audio files, I noticed there exist a small probalilty (one charpter out of total 20 charpters) that all the Kanji characters have been Seems worthy futher investigation. tiny&small model medium&large model [00:00.000 --> 00:04.000] チャプターナインティーン |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
It's similar to the "punctuation mode" and "no-punctuation mode" observed in #194, where the model can sample a certain style of writing and continue using it since the subsequent transcriptions are conditioned on the previous outputs. I think the model was more likely to go with this "no-kanji mode" given the tone and the content of the story which would be targeted to children. As mentioned in the comments in #194, you could supply a hypothetical sentence that could have come before the audio as the initial prompt, such as, |
Beta Was this translation helpful? Give feedback.
It's similar to the "punctuation mode" and "no-punctuation mode" observed in #194, where the model can sample a certain style of writing and continue using it since the subsequent transcriptions are conditioned on the previous outputs.
I think the model was more likely to go with this "no-kanji mode" given the tone and the content of the story which would be targeted to children.
As mentioned in the comments in #194, you could supply a hypothetical sentence that could have come before the audio as the initial prompt, such as,
--initial_prompt "次はヒロが見た変な夢の物語です。"
to nudge the model to output in the style you want.