Reviewing Language Reactor, Migaku, jidoujisho, etc. as an alternative to Subs2SRS

[emk]

What I'm looking for. Years ago, I used Subs2SRS and wrote substudy to work on listening comprehension in French and Spanish. I had been inspired by Sprachprofi's remarkable success with Japanese, and I got less remarkable but still awesome results from doing the same thing in Spanish. (I have accepted that Sprachprofi is about 8 times better at learning languages than I am. )

Essentially, the core of this approach is that audio cards made from native audio will get burned into your brain as earworms. And you'll train your brain to understand actual native audio (as long as it's reasonably clear). And remarkably, I can still understand many of my Anki Spanish cards after 5-8 years of total neglect. It's as burned in my brain as a popular song from high school! But the drawbacks are legion: You need accurate L2 subtitles, you need to be able to convert a TV episode or movie into video file, and you need to be pretty good with various geeky tools. And you need to have good Anki card formats and deletion habits.

Trying Language Reactor. Language Reactor seems to be frequently mentioned as a good tool for studying YouTube and Netflix videos. Let's try it out.

First, here's a French video that's still a good exercise for me. The woman in the video speaks rapidly and idiomatically, and I don't actually get everything on the first try. (At least not without a warmup.) Here's what it looks like:

langreactor-overview.png

I can set my vocabulary size, and Language Reactor will helpfully mark words I might not know. I can play the video, look at the French and English subtitles, and mark either words or phrases to be studied later. Note that the YouTube captions contain as many errors as usual:

langreactor-bad-caption.png

That caption should almost certainly be j'avais peut-être envie de passer le concours d'avocat!

There's also a browser plugin, which works with Netflix. In the video I tested, the Netflix captions seem to be more accurate than YouTube captions. But the translations are a very loose match at best.

langreactor-netflix.jpg

The flashcard support. I was optimistic I was going to like this. But my hopes were dashed quickly. Yes, your marked phrases will be automatically turned into flashcards, which look nice enough:

langreactor-flashcard.png

But the flashcards don't use the original audio! Instead, they use synthesized audio clips, which are based on the (often incorrect) YouTube captions. Compared to Subs2SRS, this is nearly useless. If you are going to use it, at least use it with Netflix content, which seems to have professionally transcribed captions.

And yes, there's an Anki export feature, but it uses the same synthesized audio.

How to make good Subs2SRS flashcards. Here's an example of how to do this right (at least in my opinion):

substudy-anki-card.png

There are several things worth mentioning here:

1. The front of the card contains a still image (for context) and the original audio.
2. The audio clip is padded with 1.5 seconds of sound before and after the subtitle's display time. This provides a bit more context, and it also helps salvage cards where the subtitle timing was off. 1.5 seconds may seem like too much, but once I started doing this, it vastly reduced the number of ruined and frustrating cards.
3. The back of the card contains bilingual subtitles. These include the preceding and following subtitles, too, because oftentimes the subtitle won't make sense on its own.

Conclusion: Language Reactor is a pretty good tool for watching Netflix! And it can be useful for YouTube videos, provided the captions aren't too bad. I like the vocabulary size estimator and the ability to mark words and phrases. This could be a very useful tool around B1 and B2, and it could be used in a variety of flexible ways.

But unfortunately, both the built-in flashcard tool and the Anki export feature use text-to-speech audio, and they appear to miss a variety of important tweaks that make good Subs2SRS cards amazingly useful. So this is not yet my dream tool. Perhaps one day!

[kundalini]

There's a service called Migaku that, I believe, is similar to Language Reactor, but does have the ability to create Anki cards with the original audio. I've never used it, so I don't know how well it works.

https://www.youtube.com/watch?v=Ong1H7BpGyU

[Greasaias]

Myself I used anime Cards MPV Script
https://anacreondjt.gitlab.io/docs/mpvscript/
which only works with Yomichan afaik - so no French support, I'm afraid.
The process is:
1. Add a card to Anki using Yomichan.
2. Go to mpv and click the add the chosen phrase to the flashcard. (The line saved in your clipboard, the corresponding audio AND a still of the scene will then be added to the recently created card.)

The setup was fairly simple - I say this as a Windows noob.

There have been implementations of the idea for other languages.
Aside from the aforementioned Migaku there is mpvacious.
However, I haven't used any of them.

Padding: maybe you could try adjusting the script?
Bilingual subtitles: no idea.

[emk]

There's a service called Migaku that, I believe, is similar to Language Reactor, but does have the ability to create Anki cards with the original audio. I've never used it, so I don't know how well it works.

Ah, thank you for the suggestion! Let's take a look.

Migaku appears to be primarily a browser plugin. It's a bit more fiddly to set up than Language Reactor. I needed to mess with my Chrome extension settings to show the icon in the address bar. Once you install it and get the icon to show, it looks like this:

migaku-config-cropped.png

Once installed, it's a bit more discreet than Language Reactor. It doesn't provide big lists of bilingual subtitles. Instead, it displays subtitles near the middle of the image, and allows you to click on words.

migaku-popup.jpg

It also works with Netflix:

migaku-netflix.jpg

In many ways, Migaku is a lot clunkier than Language Reactor. I got logged out accidentally. It refuses to work unless you keep a second browser window open. If you want to set an approximate vocabulary size for an intermediate language, Language Reactor has a fantastic tool for this, but Migaku requires you to go find a word list somewhere and paste it into a settings dialog somewhere.

But where Migaku shines is the card creator:

migaku-card-creator.jpg

The card creator has some great features:

• It captures real audio and video, even from Netflix.
• It automatically adds dictionary definitions, and a second example sentence.
• There's an "explain this word in context using ChatGPT" feature that's actually pretty good. I'm sure it hallucinates sometimes, but in a handful of tests, it gave good explanations.
• You have a choice of different card formats. They all contain the same data, but the what goes on the front can vary.

However, this is still missing some low-hanging fruit:

• It doesn't include the preceding and following subtitles. In my test, this was especially harsh with YouTube videos, where individual subtitles are usually short, and they may make little sense without more context.
• It doesn't seem to pad the captured audio. Again, adding 1.5 seconds on either side allows a much higher "yield rate" when making cards.
• It puts the captured image on the back of the card, not the front. This might be a matter of preference, but when you're reviewing 1,000+ cards with dialog, getting some indication of context is really helpful.

An older version of Migaku supported Anki export. The current version does not seem to? You can only review cards in a phone app.

The phone app (for flashcard reviews). This is fairly reasonable. Here are some screenshots.

migaku-card-front.png

migaku-card-back-2.png

migaku-card-back-1.png

In the third screenshot, you can see an example explanation added by ChatGPT. Very useful, and it's just a single click in the card creator.

Which of these tools should you use? This depends on what you want.

• Language Reactor is more polished. Migaku is more "mature", in the sense that it seems to have been written by people who have used it a lot for serious learning. But it has a lot of rough edges, too.
• Language Reactor is better for "reviewing" a video or episode intensively. In particular, it likes to show you big lists of surrounding subtitles and makes it easier to jump around. Migaku focus on watching videos, and occasionally capturing something interesting.
• Migaku is hands-down better if you're making flashcards. It uses real audio, not synthesized voices reading incorrectlt transcribed YouTube captions. It captures images. It can use ChatGPT and other tools to autofill any part of the card (and it does this well).
• Language Reactor has lists of YouTube videos, which are sort of divided up by difficulty. This is nice if you're just looking for something to watch.
• Language Reactor also has a great tool for estimating the size of your starting vocabulary, so you don't need to manually mark 5,000 words as "known," or need to manually paste in a frequency list.

I am actually tempted to try some Migaku with Netflix for Spanish.

Are there other tools in this space that people have heard a lot about?

[Amandine]

I’ve been using Migaku recently and for what it’s worth, after a period of inactivity while they were working away in the background, the team has become very active again with updates, YouTube videos on various things, plans for future features etc. I got very quick assistance from multiple people when I posted a support question in their Discord recently. So I’m hopeful it will continue to improve in the nearish future

[emk]

Thank you to everyone who's providing their experience with these tools, and who's suggesting similar tools!

Since I can't sleep this morning, some further observations about these tools...

Video sources: DVDs versus streaming. Subs2SRS requires video files in a computer readable format, and subtitles in SRT format. In practice, getting the video files requires buying DVDs from another country, having a region-unlocked DVD drive, and having software that can read the DVD to a video file. None of this is simple.

Finding SRT files requires either digging around on the Internet, or OCRing "Vobsub" subtitle tracks from a DVD. I was working on a tool to make OCR totally automatic, but I put it on the back burner for years because it involved a ton of work for no money. In practice, subtitle files tend to be annoying to find, inaccurate and misaligned. Basically, this is only feasible if you're starting a language, and if you expect to spend 5+ hours of studying per individual episode. And you need to be moderately geeky.

The new generation of tools has taken a completely different approach: They connect to online streaming video from YouTube, Netflix or Disney Plus, and they use the already existing video and subs. This has a bunch of nice properties:

1. Netflix and Disney Plus offer a wide range of content, in multiple languages, without any region-locking of specific languages. This is the most practical way by far for most US learners to watch things in their L2.
2. Netflix actually has L2 audio with (mostly?) matching L2 subs for some of the things I've looked at. This is a miracle to find for DVDs in many languages.
3. YouTube subs are pretty dodgy, but they're just good enough to be useable. There are lots of transcription errors, so you should never use text-to-speech to turn them into SRS cards—this is dangerously bad for language learning, because it will create unnatural, grammatically incorrect audio and then burn it deeply into your brain. This is a huge problem with Language Reactor.

Overall, I think that the modern tools make the right call here. They will reach the greatest number of users, and they will integrate the most naturally into extensive watching.

Good card design & good SRS tools. What I'm looking for here is:

• Easy card formats. In my personal experience, it's almost impossible to make an Anki card with native content too easy. Much of the benefit from audio and sentence cards seems to come from just seeing them again. It's nearly impossible to be too generous with hints, too. Put more stuff on the fronts of cards.
• Easy deletion. In general, Anki is vastly more pleasurable if you delete any card that makes you say "ugh, not this card again." This turns Anki from a chore into fun. Besides, if you can make cards from native content with a few clicks, cards become super cheap. Easy come, easy go. If you actually need to know a word, it will appear again soon enough.
• User-friendly UI. Since it's OK to throw out cards, and since we want the experience to be easy and pleasant, we probably don't need quite so many buttons to score a card. We could probably get away with "More often", "Less often", and "Never again", or at least auto-deletion of cards the user keeps failing.
• High "yield" when making cards from media. Seriously, I'm trying to make YouTube cards with Migaku here, and most of them are useless, because they're cut in weird places. This could be fixed! As I keep suggesting, please include 1.5 seconds of audio padding and the dialog lines before and after the current one! I'm going to keep beating this drum, because it would allow me to capture many more good cards with Migaku.
• Good card-making tools which make it easier to add context. Migaku is the best I've seen.
• An effortless path from extensive viewing to card reviews, without complicated importing or syncing.

Also, I have two more speculative ideas:

• Some kind of cloze integration or "active wave", maybe? I really want to run some experiments with this, but I suspect that taking familiar audio cards, and "flipping" them after a couple of weeks to create cloze cards, might offer some easy active practice. It's basically the Assimil active wave versus the passive wave, but the student can rely more heavily on audio memory. The goal here would be to trigger active recall of audio "chunks" in a format that helps the student prepare for conversation.
• Shorter card lifetimes. Weirdly, most of the benefit of a card seems to be achieved during the first 20-30 day gap between reviews. The reviews earlier than that seem to be the "learning" reviews. But when the card disappears for 20-30 days and then pops back up again, it often becomes dramatically easier. Challenging audio often becomes obvious. I suspect this is some form of memory consolidation or a human analog of grokking. But once this point is passed, I think it's safe to chuck the card. We don't need an Anki deck with years-old cards for this—we can safely purge audio cards older than 6-8 weeks.

This suggests that we might ideally want a tool other than Anki. But so far, all the Android alternatives to Anki are noticeably worse, particularly including the clunky Migaku app.

Anyway, I hope someone finds these reviews and ramblings useful. But when I'm sleep deprived, I'm never sure!

[jeffers]

Language Reactor is better for "reviewing" a video or episode intensively. In particular, it likes to show you big lists of surrounding subtitles and makes it easier to jump around. Migaku focus on watching videos, and occasionally capturing something interesting.

Forgetting using these tools for SRS purposes, I find Language Reactor excellent for just watching videos, which is really what it was designed for. You can set it to show your TL and your own language one above the other, and you can pause and rewind with a couple quick key taps. Even more interesting to me, you can set it to automatically pause at the end of each subtitle sentence. At first this can be jarring, but it can also be helpful for just working intensively through a section of an episode or film. In addition, you can export the subtitle to either a printable HTML file, or an Excel file, with options to include translation in parallel and timestamps. So far this is only really viable in Netflix since the subtitles are human made, because on YouTube it generates subtitles which it cannot do for languages such as Hindi.

Another feature of Language Reactor is that you can right click articles from the web and import them into Language Reactor. Within LR it shows known words in green, and you can click on words to check the dictionary. Unfortunately, you have to pay to add more words to your known list, but the ability to load up an article and produce a list of unkown words might be worth it for me. It will also read the article to you in a computerized voice. Although this voice has its limitations, this is a great feature when it comes to random news articles in more lesson studied languages on the internet!

Finally, one more advantage of LR is that it supports more languages than Migaku. I've not tested it with other lesser studied languages, but the fact that it supports Hindi makes it the winner for me by default, despite some features of Migaku that look excellent.



Images of the LR reader in action:

Screenshot 2024-03-04 212053-2.png

Screenshot 2024-03-04 212218-2.png

[emk]

Finally, one more advantage of LR is that it supports more languages than Migaku. I've not tested it with other lesser studied languages, but the fact that it supports Hindi makes it the winner for me by default, despite some features of Migaku that look excellent.

Thank you for the review of those features! That's really helpful.

I do want emphasize some key differences between the projects. Language Reactor has some great things going for it:

• It's fast and polished.
• It's really great when you want lots of bilingual text shown in two columns.
• It's excellent when you want to jump around a lot.
• It apparently supports both videos and text, as you show!
• It has some recommendations for things to watch, loosely grouped by level.

But everything related to audio cards is fatally flawed, in my opinion, by the use of synthesized audio voices. These voices have a number of problems:

• The voices read the grammatically incorrect YouTube transcriptions. This is absolutely awful, because you will be taking grammarically incorrect phrases like j'avais peut-être *vivre passer le concours d'avocat and burning them into your brain like a song lyric from high school. I am a little bit superstitious of SRS—it's powerful enough to change my fundamental understanding of a language. I only use it with actual native content if there's any choice.
• The voices are probably too slow and clear even for A1 students. If you're an absolute beginner using Subs2SRS-style cards, you should be able to handle clear natural dialog like the animated version of Avatar. Yeah, you won't be able to handle C1-level audio like Y tu mamá también, but you should be able to tackle anything that a B2-level student can watch extensively.
• Natural voices taken from interesting media are more entertaining than synthesized ones, and they carry more context. You can tell the speaker's age, level of formality, emotional state, etc.

Migaku, on the other hand, just feels clunky, in many little ways. But if you're using it to make audio cards, it has lots of advantages:

• Natural audio.
• Image capture.
• Support for ChatGPT to write (mostly) accurate explanations of how a word is used on a card. This isn't always 100% spot on, but it actually seems better than the dictionaries that both products support, at least for French and Spanish. The dictionaries mix up parts of speech, and have many different definitions for the word. ChatGPT explains how the word is used in this particular context.

Migaku also has some advantages for watching extensively:

• The Migaku interface keeps subtitles just below the center of vision, which is really neat.
• Migaku can hide subtitles until you ask for them, and it has a number of hot keys for tasks while viewing.

So my recommendation is this: If you want to do most of your studying in the video player, go for Language Reactor. If you want to watch video extensively and turn parts of it into SRS cards, try Migaku. The use of synthesized audio in Language Reactor is a deal-breaker, in my opinion, for people looking to generate audio cards. And if you're looking to extensively watch two hours of Netflix and make, oh, 20-30 SRS cards along the way, then Migaku can mostly hide itself until you want to capture a card.

And if anyone knows of any other tools in this niche that seem promising, please let me know! I'm interested in reviewing them, too. Thank you to everyone contributing information and reviews!

[emk]

Aside from the aforementioned Migaku there is mpvacious.
However, I haven't used any of them.

Padding: maybe you could try adjusting the script?

Thank you for the pointer!

mpvacious is a plugin for the open source MPV video player. This looks like it's aimed at working with locally available videos and subtitles on a computer. So if you have a carefully-collected set of L2 DVDs that you've copied to a local drive and matched up with subtitle files, this looks like a promising choice.

But I suspect that the average student is better served with a Netflix account and a browser plugin these days. Netflix doesn't require region-free DVD drives, which were a big problem for US users who wanted to play foreign DVDs. And for people studying major languages, there are quite a few good Netflix series with good dubs and (at least in the cases I glanced at) accurate subs. And of course, any browser plugin that works with Netflix works with YouTube. So I expect that's why all the new tools focus on browser plugins and streaming media.

[jeffers]

I've been seeing adverts on YouTube for something called Lingopie, with the tag-line, "Binge-watch your way to fluency". Looking at the website it seems to offer nothing more than clickable subtitles. One intriguing factor is that it claims to have "The best TV from France" with a list of channels I've (mostly) never heard of, so it might be that they give access to content unavailable without a VPN.

I've just signed up for the 7 day free trial, but didn't go the whole way to choosing a subscription, and along the way they also offered me "free group lessons". So it looks like they're trying to be all things to all learners. Another website mentions that Lingopie has "video-based flashcards", but that could mean so many different things.

I found a decently thorough review here: https://testprepinsight.com/reviews/lingopie-review/
Some interesting features they mentioned:

• Toggle back and forth between native and English subtitles or even show both at the same time!! (wow)
• Loop setting, which allows you to constantly repeat a section. Actually a pretty good idea. It would be even better if you could save your loops for later
• Playback speed (wow)
• You can repeat what a character has said and get your pronunciation rated by speech recognition technology (potentially good)
• Grammar: words are colour coded by part of speech, and when you click on a word it give grammatical information as well as dictionary definitions
• Flashcards are created automatically if you click on a word. The review says, "these are interactive, video-based flashcards that use the exact clip from the show or movie where the word or phrase was said, which really helps with recall."

For negative points, the review complained that the subtitles are not always correct. I think they meant the translation subtitles, and I think this means it is creating them with machine translation.

The last point is that Lingopie is only available for 8 languages: Spanish, French, German, Italian, Portuguese, Japanese, Korean, and Russian, which rules it out for me.

EDIT: I thought I should maybe point out that where I wrote "wow" after a feature I was being sarcastic. These features are pretty common.