Anyone can make a github account, copy a colab, modify it, host it and then share it.
Mei2 is the one who made those 2, but he's likely using something different personally and life gets in the way so he may not be aware yet that they are broken/ can't dedicate enough time to fix them.
It hasn't been that long so saying they're not maintained is premature.
I have created two colab notebooks that use WhisperTime stamped. As WhisperX, Stable-ts, and Faster Whisper, all three are currently broken on Google Colab, this model is what is quite Good. It uses standard Whisper Models and is not as fast as Faster Whisper, but it is very Accurate in time stamping. I have Tried to Remove Hallucination and Garbage, but the output File Still has few hallucinations and Garbage.
Note: It only gives Japanese subtitle file, and there is no translation funcitonality.
I mainly Use Google Gemini 3 pro, for translation and removing the hallucination and the garbage that is left. It is very good in understanding context and giving quite accurate translations. However, you can sometime struggle with content block but just tweak your prompt a bit or use "continue from where you left off" prompt to make it complete the translations.
DTW Transcription Colab Notebook
This second Notebook, is a bit more advanced. It use models to enhance the audio before it is fed to whisper. In my testing, it give more accurate timestamps, more accurate dialogues and also it even gives more dialogues that might get overlooked with the normal whisper translation. This noteboook also gives you much more control over how you want your output.
Demucs-Bs-Roformer-DTW-Transcription
Note: BS-Roofer takes quite a lot of time but makes output better (it takes around 15 minutes for 30 minute audio), use can select Demucs that is very fast and only gives boost to the output quality (not as much as BS-Roofer), choice is yours.
Tip: Although i have given the option to use silero VAD, but don't use it, as you might get a bit less hallucination but time stamps might get less accurate. Instead downgrade to large-v2, that will have less hallucination
Finally, i was able to fix Whisper with VAD Pro, you just need to use ffmpeg to feed the audio. I am not sharing it, but you can easily use google gemini or chat gpt to fix the code, just feed them the code and error and they will give the updated code, paste it back into notebook and it will work.
My colabs might not be best, but its my first try at this. I am actually consider myself more leaned towards time-stamp accuracy and more dialogue being transcripted so this workflow works for me, as i use it for plot heavy videos. But it Takes more time.
I usually convert the video to mp3, cut it into 30 minute segments, translate them with gemini 3 pro with context, then merge them together using subtitle edit. My way is far from simple or perfect, but it works for me now.
Installation Report:")
