Tested for Japanese. No problems so far, except sometimes repeating the desired number of times didn't work (mobile). Seems to work now.. But looping infinitely produces only three repetitions.
Really good UI with little friction: easily hone in on sentences, easily move on or jump ahead, see vocabulary, create Anki deck. Took a while to discover loop settings, but it's a good choice.
Only now discovered the "custom span loop mode". Great! I was about to ask for it!
AI mode is unobtrusive and helpful.
At last, found something that could need a touch up.. The starter deck from the example story is a bit nonsensical. It features words like URL and site from the Librivox intro. お is "translated" as "honorific" which is kind of true, but it's only a marker. A beginner might not know this. たち shows as answer "plural marker", there it worked.
Integrating a flashcard app is no small feat. Impressive. I wonder what algorithm was used. Does it scale?
OT but is any work anywhere being done with Japanese pronunciation problem?
Japanese language are often described as using multiple type of alphabets - kanji, kana, numbers, and English alphabets sometimes - and pronunciations of especially kanji is not very well constrained, creating tons of homophones and homographs, e.g. "koushou" shared across more than 20 words, and the character for "life" said to be involved in more than 150 differently read parts of words.
Even OT but Unicode code space used for Japanese Kanji is famously shared with Chinese Hanzi, leading to ambiguities.
This situation is causing AI-based TTS(and also image generators) trained directly on Unicode text to go weird on kanji, even for simple ones as "tomorrow". Classical pre-LLM Japanese TTS avoid this by operating on generated or manually specified pronunciations, skipping kanji altogether, which do occasionally lead to wrong readings, but won't lead to sound generation code creating butchered middle-of-road sounds.
It doesn't seem like most or any of AI TTS tackle this problem, but I'm not in that field. Do anyone know the statuses on it?
I don't know what resolution or display you built this on, but a heads up the initial impression on my 4K monitor is that everything is incredibly tiny.
To be honest I haven't tested it on a 4K monitor yet, so I am not surprised. There are two controls above the transcript that change the font size and the line spacing, which should help a bit for now. Something to fix, thanks!
- How do you transform and enrich the data? How does your pipeline look?
- What are your key challenges?
- Which tools do you use? What is your 'stack'? (Stanze, wordfreq, Whisper, wn, ...)
Background: I am currently building a multi-lang vocabulary hub for language learning. The goal is to match core words/lemmas to their senses/concepts, and then be able to generate multi-language flash cards.
I am still stuck on the sense alignment and fingerprinting (example: should 'to shop', 'einkaufen', ' alışveriş yapmak' and 'go shopping' point to the same concept of 'shop'?), but in a later stage I want to allow user-submission and data enrichment for IPA, pictograms [1] and audio.
Use-case (the dream): I come back from language class, I input new vocab and I output new Anki cards that work across all my fluent languages.
Currently, I mostly find myself knee-deep in problems of linguistics, NLP, Python and getting an LLM to do exactly what I want. At the same time it is a super fun project, and really makes me feel the joy of programming again. LLMs are magic, time just flies by, and all the random projects I always wanted to do suddenly materialize.
For coding, I mostly use free Gemini and some deepseek-v4-flash via openrouter to keep a tight oversight and understand the problem space. Maybe this slows me down, but agentic code jsut does not align with me. Overall, I haven't spent more than 2 € in total.
So far, surprisingly, the biggest problem is the lack of high-quality, free input data (example: English has the Oxford 5000 words as core vocabulary, but it is difficult to find the same for e.g. Turkish).
2nd place is the lack of high-quality synsets/wordnets (cross-language is mostly incomplete), and the 3rd place is getting LLMs to reliable play to their strength (on paper, a LLM is the perfect tool to provide multi-lang sense equivalents)
I plan to do a full writeup sometimes, but first I need it to work :)
Thanks!
As far as I understand your idea is to starts from the word and pulls examples from some huge data source. My approach is the other way round: I start from a source (the audio that you want to learn), and the tool extracts only the words that appear in it, with their meaning in that context. I think that hugely simplifies the implementation, and it is more useful for learners. They learn the meaning in a particular context.
As for the stack: STT with Soniox (word level timestamps), then spaCy for segmentation, POS and lemmas, then AI enrichment, correcting the lemma when spaCy is wrong. Some languages have no spaCy model at all and others are unreliable. I am trying to do spaCy thing in LLM then. Plus some extra magic for Japanese and Chinese.
Awesome, and yes, totally makes sense -- you are more learner-centric that way.
Having the full sentence context is actually one of the things I have been thinking about a lot -- this helps both the learner as well as the POS detection in Stanza. I always decided against, because I wanted to build agnostic flash-cards.
However, as your approach allows on-the-fly generation of flash cards, you always stay close to the learner progress. I could (e.g.) pick some Gutenberg fairy tales, allow the learner to read them in their target language and provide bi- and omni-directional translations across all languages. Creating flash cards from the source material keeps the learner in progress (context), allows to learn new words step-by-step (discovery), as well as providing a fun learning experience and measurable progress. Similarly, instead of fairy tales, we could use some series in combination with its subtitles. This allows video-progress. Awesome x2!
Sidenote: The awesome part about HN is that I get to chat with like-minded people and directly grasp some new inspiration. Probably I ought to visit some in-person hacker spaces :)
If you don’t mind sharing, how much does that cost you to integrate the translation API, and the text to speech API you’re using? Just curious as I’ve been thinking about doing something in that area (not anki or translations, but also language learning related).
No TTS at all in my app :) that was a deliberate choice, only STT. I experimented with many STT options, even self hosting Whisper, but ended up with Soniox. A bit expensive, but reliable. For the AI enrichment I went with Gemini Flash. I also tried Gemma 31B, which is really cheap and surprisingly good, on par with Gemini Flash, but extremely slow everywhere I tried. So you can make your own calculations :) And thanks for the congrats!
I like the structure of their privacy policy page [0] and how it appears that they are not data-greedy.
And the site itself is a great idea and implementation, though the font size and family of the ui (not of the actual playback area) has a lot of room for improvement, but those are just minor changes.
Thanks, just to clarify "they" is actually only me :) I'm a contractor and run this through my own company. I try to collect as little as possible. And you're right about the UI fonts it's clearly something I need to fix ASAP. Appreciate the feedback!
This is awesome! I’ll be lurking for new data sources. I’m working on a self-hosted language app more focused around cloze and sentence mining into Anki. I love seeing more stuff happening in this space
Thanks! I am glad you like it! I essentially mine the source audio, and all examples have cloze style gaps (blurring, in my case) that are revealed on the back of the card. I also beep the word in the sentence when you try to play it on the front card in built-in SRS system. Unfortunately that is not implemented in the Anki export, but it is technically possible.
Thanks! Chinese and Japanese as source languages are still experimental, I did my best to support them but I have to rely on people who actually know the language and this kind of feedback is really useful. I'll look into adding traditional characters and fixing the pinyin.
Although there are some clashes that it does not handle, e.g. 隻 and 只 are both 只 in simplified, you just have to know which one it is from context, but the extension fails to convert to 隻 where appropriate.
Thanks, really useful extension link. Proper traditional support probably needs a context aware layer, not a plain lookup. I will experiment with additional LLM enrichment. Appreciate you digging into this!
I also built a tool to help me study Spanish. I really like the idea of shadowing, so I built a tool that lets you take any YouTube video and generate a sentence-by-sentence exercise to help you repeat the speaker's phrases.
For segmentation and POS I rely on spaCy zh_core_web_sm, pinyin from pypinyin library. Also the small correction level on top. But I am not a Chinese language expert to judge if it really works and I'll rely on feedback from the users to improve it.
Yes, the transcriber API I use (Soniox) actually supports more than 60 languages. I just didn't have any automated testing for them. The way I tested was to find audio with a reliable reference transcription and put it through my pipeline. Then compare the results. Also some languages don't have reliable libraries to get part of speech and lemmas, something that flashcard needs.
Thanks! Yes, it's getting better for Greek but still not on par with other languages. I completed the only 2 Greek levels on Duolingo and they are really boring compared to the German one I am doing now. Easy Greek is a bit above my level, and the number of YouTubers in Greek is tiny compared to German.
Very nice work. I'm going for a different thing, but my audio2anki tool [1] is about as streamlined as I could make it to turn a YouTube URL I want to learn into a stack of Anki flashcards, purely locally.
This is really cool, just as I'm starting to get towards the back end of the Kaishi 1.5k deck so this will be perfect for my Japanese studies. Thanks for sharing.
This is awesome.
Tested for Japanese. No problems so far, except sometimes repeating the desired number of times didn't work (mobile). Seems to work now.. But looping infinitely produces only three repetitions.
Really good UI with little friction: easily hone in on sentences, easily move on or jump ahead, see vocabulary, create Anki deck. Took a while to discover loop settings, but it's a good choice.
Only now discovered the "custom span loop mode". Great! I was about to ask for it!
AI mode is unobtrusive and helpful.
At last, found something that could need a touch up.. The starter deck from the example story is a bit nonsensical. It features words like URL and site from the Librivox intro. お is "translated" as "honorific" which is kind of true, but it's only a marker. A beginner might not know this. たち shows as answer "plural marker", there it worked. Integrating a flashcard app is no small feat. Impressive. I wonder what algorithm was used. Does it scale?
That's all. Thumbs up!
OT but is any work anywhere being done with Japanese pronunciation problem?
Japanese language are often described as using multiple type of alphabets - kanji, kana, numbers, and English alphabets sometimes - and pronunciations of especially kanji is not very well constrained, creating tons of homophones and homographs, e.g. "koushou" shared across more than 20 words, and the character for "life" said to be involved in more than 150 differently read parts of words.
Even OT but Unicode code space used for Japanese Kanji is famously shared with Chinese Hanzi, leading to ambiguities.
This situation is causing AI-based TTS(and also image generators) trained directly on Unicode text to go weird on kanji, even for simple ones as "tomorrow". Classical pre-LLM Japanese TTS avoid this by operating on generated or manually specified pronunciations, skipping kanji altogether, which do occasionally lead to wrong readings, but won't lead to sound generation code creating butchered middle-of-road sounds.
It doesn't seem like most or any of AI TTS tackle this problem, but I'm not in that field. Do anyone know the statuses on it?
I don't know what resolution or display you built this on, but a heads up the initial impression on my 4K monitor is that everything is incredibly tiny.
To be honest I haven't tested it on a 4K monitor yet, so I am not surprised. There are two controls above the transcript that change the font size and the line spacing, which should help a bit for now. Something to fix, thanks!
Very cool!
Are you willing to share more technical details?
- Which data sources do you ingest?
- How do you transform and enrich the data? How does your pipeline look?
- What are your key challenges?
- Which tools do you use? What is your 'stack'? (Stanze, wordfreq, Whisper, wn, ...)
Background: I am currently building a multi-lang vocabulary hub for language learning. The goal is to match core words/lemmas to their senses/concepts, and then be able to generate multi-language flash cards.
I am still stuck on the sense alignment and fingerprinting (example: should 'to shop', 'einkaufen', ' alışveriş yapmak' and 'go shopping' point to the same concept of 'shop'?), but in a later stage I want to allow user-submission and data enrichment for IPA, pictograms [1] and audio.
[1: https://arasaac.org/pictograms/search]
Use-case (the dream): I come back from language class, I input new vocab and I output new Anki cards that work across all my fluent languages.
Currently, I mostly find myself knee-deep in problems of linguistics, NLP, Python and getting an LLM to do exactly what I want. At the same time it is a super fun project, and really makes me feel the joy of programming again. LLMs are magic, time just flies by, and all the random projects I always wanted to do suddenly materialize.
For coding, I mostly use free Gemini and some deepseek-v4-flash via openrouter to keep a tight oversight and understand the problem space. Maybe this slows me down, but agentic code jsut does not align with me. Overall, I haven't spent more than 2 € in total.
So far, surprisingly, the biggest problem is the lack of high-quality, free input data (example: English has the Oxford 5000 words as core vocabulary, but it is difficult to find the same for e.g. Turkish).
2nd place is the lack of high-quality synsets/wordnets (cross-language is mostly incomplete), and the 3rd place is getting LLMs to reliable play to their strength (on paper, a LLM is the perfect tool to provide multi-lang sense equivalents)
I plan to do a full writeup sometimes, but first I need it to work :)
Thanks! As far as I understand your idea is to starts from the word and pulls examples from some huge data source. My approach is the other way round: I start from a source (the audio that you want to learn), and the tool extracts only the words that appear in it, with their meaning in that context. I think that hugely simplifies the implementation, and it is more useful for learners. They learn the meaning in a particular context.
As for the stack: STT with Soniox (word level timestamps), then spaCy for segmentation, POS and lemmas, then AI enrichment, correcting the lemma when spaCy is wrong. Some languages have no spaCy model at all and others are unreliable. I am trying to do spaCy thing in LLM then. Plus some extra magic for Japanese and Chinese.
Awesome, and yes, totally makes sense -- you are more learner-centric that way.
Having the full sentence context is actually one of the things I have been thinking about a lot -- this helps both the learner as well as the POS detection in Stanza. I always decided against, because I wanted to build agnostic flash-cards.
However, as your approach allows on-the-fly generation of flash cards, you always stay close to the learner progress. I could (e.g.) pick some Gutenberg fairy tales, allow the learner to read them in their target language and provide bi- and omni-directional translations across all languages. Creating flash cards from the source material keeps the learner in progress (context), allows to learn new words step-by-step (discovery), as well as providing a fun learning experience and measurable progress. Similarly, instead of fairy tales, we could use some series in combination with its subtitles. This allows video-progress. Awesome x2!
Sidenote: The awesome part about HN is that I get to chat with like-minded people and directly grasp some new inspiration. Probably I ought to visit some in-person hacker spaces :)
If you don’t mind sharing, how much does that cost you to integrate the translation API, and the text to speech API you’re using? Just curious as I’ve been thinking about doing something in that area (not anki or translations, but also language learning related).
Great project, and congrats for launching :)
No TTS at all in my app :) that was a deliberate choice, only STT. I experimented with many STT options, even self hosting Whisper, but ended up with Soniox. A bit expensive, but reliable. For the AI enrichment I went with Gemini Flash. I also tried Gemma 31B, which is really cheap and surprisingly good, on par with Gemini Flash, but extremely slow everywhere I tried. So you can make your own calculations :) And thanks for the congrats!
I like the structure of their privacy policy page [0] and how it appears that they are not data-greedy.
And the site itself is a great idea and implementation, though the font size and family of the ui (not of the actual playback area) has a lot of room for improvement, but those are just minor changes.
[0] https://lingochunk.com/privacy
Thanks, just to clarify "they" is actually only me :) I'm a contractor and run this through my own company. I try to collect as little as possible. And you're right about the UI fonts it's clearly something I need to fix ASAP. Appreciate the feedback!
This is awesome! I’ll be lurking for new data sources. I’m working on a self-hosted language app more focused around cloze and sentence mining into Anki. I love seeing more stuff happening in this space
Thanks! I am glad you like it! I essentially mine the source audio, and all examples have cloze style gaps (blurring, in my case) that are revealed on the back of the card. I also beep the word in the sentence when you try to play it on the front card in built-in SRS system. Unfortunately that is not implemented in the Anki export, but it is technically possible.
Is it possible to add traditional characters for mandarin?
Also the pinyin for 誰/谁 is coming through as shuí, whilst this character has two pronounciations, I believe shéi is the more common one.
Thanks! Chinese and Japanese as source languages are still experimental, I did my best to support them but I have to rely on people who actually know the language and this kind of feedback is really useful. I'll look into adding traditional characters and fixing the pinyin.
No worries, I appreciate the effort. I did go back and listen and they are indeed pronouncing sheí in the audio too.
I use a firefox extension to convert simplified to traditional, looks like it's open source so that may be of some use to you: https://github.com/tongwentang/tongwentang-extension.
Although there are some clashes that it does not handle, e.g. 隻 and 只 are both 只 in simplified, you just have to know which one it is from context, but the extension fails to convert to 隻 where appropriate.
Thanks, really useful extension link. Proper traditional support probably needs a context aware layer, not a plain lookup. I will experiment with additional LLM enrichment. Appreciate you digging into this!
I also built a tool to help me study Spanish. I really like the idea of shadowing, so I built a tool that lets you take any YouTube video and generate a sentence-by-sentence exercise to help you repeat the speaker's phrases.
https://talkhabit.com/shadow Or example, of one exercise: https://talkhabit.com/shadow?videoUrl=https%3A%2F%2Fwww.yout...
Stuff I need to work on: - It only works with videos that have auto-generated captions - It works best with monologue videos
What are you doing for Chinese word segmentation/pinyin?
For segmentation and POS I rely on spaCy zh_core_web_sm, pinyin from pypinyin library. Also the small correction level on top. But I am not a Chinese language expert to judge if it really works and I'll rely on feedback from the users to improve it.
Just tried it with an unsupported language and it still worked I set it to Chinese and inputted the audio. Still got correct results.
Yes, the transcriber API I use (Soniox) actually supports more than 60 languages. I just didn't have any automated testing for them. The way I tested was to find audio with a reliable reference transcription and put it through my pipeline. Then compare the results. Also some languages don't have reliable libraries to get part of speech and lemmas, something that flashcard needs.
Very cool! I'm also learning Greek and it's amazing how many resources are becoming available.
Thanks! Yes, it's getting better for Greek but still not on par with other languages. I completed the only 2 Greek levels on Duolingo and they are really boring compared to the German one I am doing now. Easy Greek is a bit above my level, and the number of YouTubers in Greek is tiny compared to German.
Very nice work. I'm going for a different thing, but my audio2anki tool [1] is about as streamlined as I could make it to turn a YouTube URL I want to learn into a stack of Anki flashcards, purely locally.
[1]: https://github.com/hiAndrewQuinn/audio2anki
This is really cool, just as I'm starting to get towards the back end of the Kaishi 1.5k deck so this will be perfect for my Japanese studies. Thanks for sharing.
Thanks, I hope it will be helpful! If anything looks off, please let me know.