For Japanese to English, the transcription alone is already pretty inaccurate (usable if you know some Japanese; but then again you already know Japanese!)
As long as you're expressive enough in English, and reverse the translation direction every now and again to double check the output then it works fine.
Machine translation for instructional or work-related texts has been "usable" for years, way before LLM emerged.
LLM-based translation has certainly made significant progress in these scenarios—GPT-4, for example, is fully capable IMHO. However, it's still not quite fast enough for real-time use, and the smaller models that can run offline still don't deliver the needed quality.
Anyway, the current state of affairs float somewhere comfortably above "broken clock" and unfortunately below "Babelfish achieved", so opinions may vary.
I can't say much about the quality of English -> Japanese translation, except that people were generally able to understand whatever came out of it.
But don't expect to be able to use it to read actual literature or, back to the topic, subtitling a TV series or a YouTube video without misunderstanding.