The Process - Part 2

After the first Whisper passthrough, now you have to sift through the entire .srt file and quality check it.

STEP 2 - QUALITY CHECK

It's usually pretty bad. Even when you use a Voice Activity Detector (VAD) with Whisper on JAVs, it'll still create subtitles for moaning, screaming, and other miscellaneous noises. These are no good.

However, as I mentioned in the previous post, the goal of the first Whisper passthrough is to get an overall idea of the title. Is there actually a story? Does the actress just walk up to another actor and start their business? Is there an interview? A montage? This first passthrough will answer all of these questions.

STEP 3 - ISOLATE AUDIO TO WHISPER

Then, I open the .srt file with SubtitleEdit, an open-source application that lets you edit subtitles alongside with a video preview and audio waveform. It also has Whisper directly implemented into the application, so you can apply Whisper to a particular segment of audio.

This is the bulk of the work. About 90% of The Process is isolating the audio in SubtitleEdit and using Whisper to transcribe them.

STEP 4 - ONLINE TRANSLATION TOOLS

That's right. I don't use Whisper to translate at this point. Instead, after Whisper has transcribed the audio into Japanese text, I paste the text into Papago, DeepL, and Google Translate. Why use three different ones? All of them have strengths and weaknesses.

PAPAGO

PROS

• Romaji pronunciation at a glance.
• Can generally understand the meaning from just Hiragana.

CONS

• Transliterates names (If you give it the kanji "Akane", it'll literally translate it as "deep red" instead of recognizing it as the name "Akane.").
• Does not understand dirty talk or slang.

DEEPL

PROS

• Translation tool based on large language models (similar to other AI tools).
• Understands dirty talk and explicit terms.

CONS

• No romaji pronunciation at a glance.
• Transliterates names.

GOOGLE TRANSLATE

PROS

• Romaji pronunciation at a glance.
• Offers suggestions when provided with only Hiragana.
• Dictionary and thesaurus with selected text.
• Has a "Jinmeiyo" dictionary (meaning it will recognize "Akane" as "Akane" and not "deep red.").
• Can translate websites.

CONS

• Sometimes understands dirty talk and slang.

As you can see, it's better to cross-reference all three translation tools to determine what exactly the segmented audio is trying to say. Sometimes one site says it better and sometimes another will say it differently.

Then, once you understand what the audio is trying to say, it's time to localize the text.

STEP 5 - LOCALIZE

This is what separates my work from the other English subtitles out there.

While most people are pretty happy with stilted, unnatural English for their subtitles (as long as they get the point across), adding another round of quality check makes a HUGE difference.

You are watching porn. You can read the line "it feels good" so many times until it starts to really bother you and you lose interest. However, if you can match the idea and context of the video with your subtitles, it'll make it much more engaging and entertaining.

Instead of saying "it's because you're so pretty", which would be a 100% accurate translation, I would translate it as "it's because you're so fucking hot", which would be 60% accurate, but 100% more erotic.

My philosophy is SEXY over ACCURACY.

Again, you are watching porn. Accuracy of translation is the least of your concern if it doesn't turn you on. It needs to be erotic, spicy, and engaging.

Of course, you need to match the level of eroticism appropriately according to the context of the scene. If the scene involves two lovers, you would probably tone down the dirty talk into something more romantic and wholesome. However, if the scene is more visceral and about raw action, you would increase the amount of explicit language and dirty talk.

And honestly, I think it makes the viewing experience much more enjoyable if you have subtitles that match the level of action on the screen. That's why I translate and localize them in this fashion. I will never go back to unnatural and uninspiring dialogue.

That's it for The Process! I do this for every title I personally work on and for every request. It's quite a lengthy process, but in my opinion, the effort is always worth it.