I used to be terrible at languages. I saw the advantage of it, but the method used at high school didn't work for me. A few years ago while commuting by train, and decided to do "something fun but useful", and downloaded DuoLingo to learn Italian. After a while, it became obvious that Duolingo would only get you started, not to intermediate or advanced levels of language comprehension (even more so now than 10 years ago).

So I was stuck again, but the idea of doing daily flashcards worked for me. After a while I realized that I had mostly learned English by using the language (first by listening and reading, and later by speaking and writing). So I decided that would be my next attempt at this silly project of learning Italian: reading or listening something every day. Along the way I discovered some YouTube channels that discuss this approach as Comprehensible Input. One of the first channels I found was Matt vs Japan, but if you're interested I would recommend the Refold channel (and after that Language Simp, Steve Kaufman, Days and Words, Luca Lampariello). Many people find one approach very effective: After learning some language basics, just read or listen something that's interesting and somewhat challenging for your level in the language, and create some flashcards daily (a new word, expression, word ordering you found surprising). Rinse and repeat. It's like a flywheel that steadily increases your level. Check out one of the channels above if you're interested.

Ok, so where does AI fit into this? I try to keep my work separate from my hobbies, but I'm still a nerd :) And there are some good use cases.

Flashcards

You can use physical flashcards, but there are advantages to using a program: You can add sound or a picture, you always have them with you (on your phone), and most importantly: optimal scheduling! A program like Anki with FSRS enabled is very good at estimating when you are about to forget something, presenting the flashcard that particular day. This allows you to review thousands of cards, only spending 10 minutes a day. Is the scheduling AI? Enough people call linear regression AI, and FSRS is a form of regression. It's a very effective and useful one.

Speech-to-text

Imagine you found a series you like, but it doesn't come with subtitles. You can use a deep learning model like Whisper by Meta to generate them for you. This is very useful if the characters are speaking a little too fast, are mumbling or have a heavy accent. The model isn't perfect, but think or it as a friend who is slightly better at the language than you, and is trying to help.

Text-to-speech

One downside of reading is that you don't know the exact pronunciation of words. Or the melody of a sentence (yes, different languages have a different melody / flow to them). So when I create flashcards from text I often use a text-to-speech engine to add audio to a flashcard. Currently Elevenlabs' commercial service is the best (and the 5 euro per month plan is sufficient), but this will probably change in the future.

Translation

It's been so long since Google Translate came out, you would almost forget it was a very impressive feat of statistical programming at the time. Currently DeepL is slightly better (especially with metaphors), but there are many alternatives. (Not AI, but very useful: wordreference.com)

ChatGPT

In this day and age it's almost a crime not to talk about LLMs. Are they useful? Yes, but not as much as you would think, mostly because of hallucinations. You can ask it to translate a sentence, or to find grammar mistakes in a text, or to write you a short story using some words you found, or provide example sentences for flashcards, etc. It's all very impressive, and it's a powerful tool to use. However when it convinces you of something very incorrect, because of hallucinations, and you find out later, it's just very annoying and demotivating. Local models are worse than the full OpenAI / Claude models.

Common Voice dataset

There are a few others, but I'm talking about datasets with spoken sentences. Why are these interesting? Well, the audio from a TTS model is fine, but natives speaking their own language is exactly what you want. Also, there are no good TTS options for smaller languages.

ASBplayer

I have to talk about this tool. It's not AI, but the combination of Anki + ASBplayer is just excellent. It's is a tool to create flashcards from video (YouTube and Netflix). When you encounter dialog with a new word, or an unexpected pronunciation, you can use this tool to: make a screenshot, record the audio, extract the target language text from the subtitles, the translation from your native language subtitles, and finally export the flashcard into Anki. This creates very high quality cards with minimal effort.

Hope this is useful to someone :)