Whisper (OpenAI) - Automatic English Subtitles for Any Film in Any Language - An Intro & Guide to Subtitling JAV

[SamKook]

Seems odd to me that you have 2 different python folder involved(if you use a virtual environment, the system python won't see installed modules on it for example and vice-versa), but I don't know how his installation works at all so that might be normal.

 

[ToastFrench]

Seems odd to me that you have 2 different python folder involved(if you use a virtual environment, the system python won't see installed modules on it for example and vice-versa), but I don't know how his installation works at all so that might be normal.

It's probably me screwing up the installation, lol. I'm really trying to understand Python and the virtual environments like venv and pyenv, but I'm just not great at this stuff.

Edit: I'm up and running!

 

[mei2]

Oops, I mixed up the threads. Ijust posted this in the other thread:

M

Post in thread 'akiba resident JAV subtitlers & subtitle talk★NOT A SUB REQUEST THREAD★'

Thursday at 8:38 PM

Just a quick update, I just did a guide for using WhisperJAV:

WhisperJAV User Guide


The guide so far includes Windows, Mac, and Linux usage. It doesn't include the colab and kaggle releases. Those notebooks are basically implementing the ensemble mode. The difference is that in Kaggle each pass runs in parallel on their dedicated GPU. In colab the two pass are orchestrated sequentially.

• mei2

• 
• 

 

[mei2]

EDIT: I fixed it: For Mac users using Brew, you might need to run:
brew install portaudio
pip install pyaudio

Thanks. I added them to the install script and to the docs.
Again, I don't have a Mac to test, so just let me know if anything pops up.

 

[ToastFrench]

I ran in "Balanced" mode, and it worked beautifully (if slowly, since it was CPU only-about 2 hours for a 2 hour video) The quality of translation is just really, really good-thank you for your hard work!

I'm now trying "Transformers" mode, which supports Metal, as I would like to give you some benchmarks between a MacMini M1 (Geekbench Metal 32K) and a MBP M1Pro (Geekbench Metal 68k)...it initially tried to just translate without Transcribing, but I think setting both the initial Transcription tab and Ensemble tab to Transformers mode has fixed it-running tests on both machines.

 

[mei2]

I ran in "Balanced" mode, and it worked beautifully (if slowly, since it was CPU only-about 2 hours for a 2 hour video) The quality of translation is just really, really good-thank you for your hard work!

I'm now trying "Transformers" mode, which supports Metal, as I would like to give you some benchmarks between a MacMini M1 (Geekbench Metal 32K) and a MBP M1Pro (Geekbench Metal 68k)...it initially tried to just translate without Transcribing, but I think setting both the initial Transcription tab and Ensemble tab to Transformers mode has fixed it-running tests on both machines.

Thanks that would be helpful.
Meanwhile, give a try to ensemble mode 2-pass. You will like the results

 

[ToastFrench]

Thanks that would be helpful.
Meanwhile, give a try to ensemble mode 2-pass. You will like the results

I'll be sure to give them all a workout!

I'm seeing this error in the logs (on both Macs), but it still seems to run-is it not that important?

/Users/Toast/venvs/whisperjav/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (7.0.1)/charset_normalizer (3.4.4) doesn't match a supported version!

 

[mei2]

I'm seeing this error in the logs (on both Macs), but it still seems to run-is it not that important? /Users/Toast/venvs/whisperjav/lib/python3.12/site-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.6.3) or chardet (7.0.1)/charset_normalizer (3.4.4) doesn't match a supported version!

Yes, that is a harmless dependency version warning

 

[ToastFrench]

Sorry for all the Mac questions!
I am trying the Metal-enabled Transformers mode:

2026-03-06 12:12:59 - whisperjav - INFO - Step 3/5: Transcribing with HF Transformers...
2026-03-06 12:12:59 - whisperjav - INFO - Transcribing full audio file...
2026-03-06 12:13:02 - whisperjav - INFO - Loading HF Transformers ASR pipeline...
2026-03-06 12:13:02 - whisperjav - INFO - Model: kotoba-tech/kotoba-whisper-bilingual-v1.0
2026-03-06 12:13:02 - whisperjav - INFO - Device: cpu
2026-03-06 12:13:02 - whisperjav - INFO - Dtype: torch.float32
2026-03-06 12:13:02 - whisperjav - INFO - Attention: sdpa
2026-03-06 12:13:02 - whisperjav - INFO - Batch: 8
`torch_dtype` is deprecated! Use `dtype` instead!
Device set to use cpu

Based on MacOS Activity Monitor, Transformers mode is using only CPU, no GPU cycles.
I have used your new guide to check, and my machine is seen as MPS, arm64, and I updated PyTorch to confirm they were default and not CUDA.
is there anything else I can do to force HuggingFace Transformers mode to use MPS?

Edit: I do not believe the CPU mode worked; the .SRT file was 0 bytes:

2026-03-06 13:31:54 - whisperjav - INFO - TransformersASR transcribing with task='translate' - output should be in English
Whisper did not predict an ending timestamp, which can happen if audio is cut off in the middle of a word. Also make sure WhisperTimeStampLogitsProcessor was used during generation.
2026-03-06 14:22:21 - whisperjav - WARNING - Translation mode was requested but output appears to be in Japanese (14437/17022 chars are Japanese). This may indicate HuggingFace translation is not working as expected.
2026-03-06 14:22:21 - whisperjav - WARNING - Sample output: 皆さん、こんにちは。今日も始まりました。アイヘンリオンのビジルトレーニング、皆さん3分間、しっかりと頑張っていきましょう。今日も頑張っていきましょう。いいですね。ちょっと待って、一緒に一緒に頑張ります...
[DONE] Transcription complete: 1 segments

 

[mei2]

Based on MacOS Activity Monitor, Transformers mode is using only CPU, no GPU cycles.
I have used your new guide to check, and my machine is seen as MPS, arm64, and I updated PyTorch to confirm they were default and not CUDA.
is there anything else I can do to force HuggingFace Transformers mode to use MPS?

Edit: I do not believe the CPU mode worked; the .SRT file was 0 bytes:

I'll look into it during the weekend. Can you do a run while you have the debug option on --the debug flag is in the adv. options tab. If you could copy all the console putput and send to me, that would be helpful.
Also, Have you tries Qwen3-ASR and anime-whisper? Do you get same issues from those?

Meanwhile, I must say the performance of transfromers so far has been disappointing --so you're not missing much

 

[ToastFrench]

I'll look into it during the weekend. Can you do a run while you have the debug option on --the debug flag is in the adv. options tab. If you could copy all the console putput and send to me, that would be helpful.
Also, Have you tries Qwen3-ASR and anime-whisper? Do you get same issues from those?

Meanwhile, I must say the performance of transfromers so far has been disappointing --so you're not missing much

I will run it tonight (I may be on a different time zone than you). I will also try to create a GitHub account(!) to put these issues/logs, if that is better.

I have tried anime-whisper as part of an ensemble, but...I am running into a similar problem with local LLMs as these users did: https://github.com/meizhong986/WhisperJAV/issues/196
So I have a couple subtitles successfully generated with it and balanced, but in Japanese.


Oh, you addressed it tonight, lol.
I did reduce Batch to 10, but that didn't help.
I will try your other suggestion to lower Scene Threshold as well to 30.

On a related note, is there a reason the Direct To English (Whisper) option is available in the basic Transcription, but is not available for Ensemble or Standalone? It might be useful for testing quickly, as it seems to provide decent results.
Of course, if I can get the local LLM thing to work, maybe that is just a superior solution.

 

[mei2]

I will run it tonight (I may be on a different time zone than you). I will also try to create a GitHub account(!) to put these issues/logs, if that is better.

I have tried anime-whisper as part of an ensemble, but...I am running into a similar problem with local LLMs as these users did: https://github.com/meizhong986/WhisperJAV/issues/196
So I have a couple subtitles successfully generated with it and balanced, but in Japanese.


Oh, you addressed it tonight, lol.
I did reduce Batch to 10, but that didn't help.
I will try your other suggestion to lower Scene Threshold as well to 30.

On a related note, is there a reason the Direct To English (Whisper) option is available in the basic Transcription, but is not available for Ensemble or Standalone? It might be useful for testing quickly, as it seems to provide decent results.
Of course, if I can get the local LLM thing to work, maybe that is just a superior solution.

I just released a beta for 1.8.7 which should fix the problem you're facing with the local LLM.