## Overview ### Immersion only works at your level The biggest jump in vocabulary in the first few months of learning a language is going from textbook and children's content to media on topics you actually enjoy, due to the large gap in domain-specific language. Immersion is supposed to close that gap, but only pays off when you already understand most of what is presented to you. Below that line, the immersion content is indistinguishable from noise. The hard part in this process for me was finding Telugu media at my level before spending time on it. ### Lack of Telugu NLP infrastructure Telugu does not have substantive NLP infrastructure to support a sentence mining pipeline. There is no popular lemmatizer, and because Telugu is agglutinative, it appends suffixes for case markers, pronouns, and tense. So, we have to account for countless combinations of inflected forms compared to English. ## Solution Ganula is a language immersion tool. At its core, is a trackable pop-up dictionary for all media formats: video, audio, and text. The pipeline turns phrases you look up into flashcards you can review using a spaced repetition system like Anki. It currently ships Telugu, but can be generalized for any language. I started building Ganula four months into learning Telugu, because I love Telugu media, especially music, and wanted to effectively learn from immersion. ### User flow View a Telugu web video or audio, or open a text page. With the Chrome extension active, Ganula renders the text as clickable tokens. These are subtitle cues that can override the native web subtitles. Hovering over a phrase defines the word with a dictionary popup. Every word tracks a known status (unknown or tracking or known or ignored), with inline color indicators, so a transcript becomes an intuitive map of what to learn next. Word pop-ups can be sent to the card creator in one click, which fills flashcard fields with the root lemma, the definition, the example sentence the word was found in, English transliterations, an audio clip from the video/audio source, and a screenshot from the video source. Cards can be saved in the app or exported to Anki to review through spaced repetition. Ganula also supports a native media player, allowing you to upload your own video, audio, and text media. ![[/media/ganula/dictionary.png|One click: resolved lemma, transliteration, layered definitions, and the card creator.]] ### Recommendation System To close the gap between introductory learning and immersion learning, Ganula features an extensive media library that features videos, podcasts, music, poems, books, lessons, and movies. The recommendation system is tailored towards a user's vocabulary level, with comprehension scores for each media item which grades the media based on the approachability of the transcript and vocabulary. ## Language Learning ### Comprehensible input Comprehensible input is a popular, well-reviewed technique in the language learning community theorizing that you acquire a language best by understanding sentences just beyond your current level, which he wrote as i+1, and the practice of mining sentences out of native media into flashcards was popularized by the online Japanese learning community. n+1 is a general rule describing a sentence where you know every word but one, allowing you to "mine" the word in context of the sentence. This also exposes a limit in the comprehension score, which is that knowing you know some percentage of the words in a video does not tell you the density of n+1 sentences. I estimate the usable threshold as somewhere around 50% when close studying and 70% to for seamless use. Ganula transcript view allows you to select all n+1 sentences at once and send them all to the card creator for batch sentence mining. The comprehension score also gives information on the number of n+1 and recommended sentences to learn. :::figure :::grid :::card{label="n+1"} :::bars ::: Word is easily mineable in context of the surrounding sentence. ::: :::card{label="n+3"} :::bars ::: Many unknown words can make it hard to anchor vocabulary to context. ::: ::: ::: ### Spaced repetition Spaced repetition works if you use it right. But the way most people build cards — the Telugu word on one side, the English counterpart on the other, and that is it — leaves you leaning on rote memorization and mnemonics. The deeper problem is that a card in that shape abstracts away the context of the word. Sourcing vocabulary from an actual corpus fixes the shape. The card carries the sentence the word was pulled from, a snapshot of the frame if it is video or the audio clip if it is not, and — instead of a counterpart — a definition in your native language. Then you are not matching one to one. You are understanding what the word actually means in the target language, the way you would use a dictionary, because words almost never have exact equivalents across languages. Context does not stop at the sentence, though. The paragraph, and the whole body of text, create it too — all of them at degrees of vicinity. You cannot jam an entire text onto a flashcard, but you can structure the vocabulary so it jogs that part of your memory and resurfaces the full context. Ganula does not do that yet. :::figure{caption="The same word, with and without its context."} :::grid :::card{label="Word-counterpart card"} :::anatomy

Front: the Telugu word
Back: one English word

::: ::: :::card{label="Corpus-sourced card"} :::anatomy

Front: the word, in the sentence it came from
Audio: the clip, cut at the cue boundaries
Frame: the grab from the video canvas
Back: a definition, not a counterpart
Reading: ISO 15919 for word and sentence

::: ::: ::: ::: ### The cold-start problem The version worth building is a card and memory system of Ganula's own that syncs with an Anki deck — ideally its repetition status too — so learners who do not use Anki, or who are on a phone, are not shut out. Most people have never used Anki. Which leads to the harder problem. n+1 is measured against a vocabulary you have declared, so the sweep is only ever as good as what Ganula knows you know; its empty state literally reads *mark more words as known first*. Learners who already have a deck can seed that: paste a word list or upload an Anki text export, five thousand words at a time. Learners who do not start from zero, and reaching a real known level in Telugu means marking something on the order of tens of thousands of words, one at a time. Even work that simple is ridiculous at that volume. Reducing that front-loading is unsolved, and it is a large part of why the tool is still in development. ## Technology ### Resolving a word nobody wrote down Telugu is agglutinative, so exact-match lookup misses most of what a learner clicks. Lookup runs as a cascade: exact match first, then the rule-based suffix stripper, then a language model guessing the lemma, and finally a generated definition if everything above missed. Latin-script tokens skip the Telugu path entirely, since code-switching into English mid-sentence is normal rather than an edge case. Underneath it the dictionary is layered — Brown, then Wiktionary, then generated entries — and results are sorted by source on every lookup rather than trusting row order out of the database, so a learner can tell which claim came from a lexicographer and which came from a model. The whole shape is provisional. It exists because there was no Telugu NLP model to call — something like spaCy, but for Telugu, that structures the data well enough to skip the guesswork. If one turns out to exist, most of this cascade stops being necessary. :::figure{caption="Every click walks down until something resolves."} :::steps 1. **Exact match**the clicked form is already an entry 2. **Rule-based suffix stripper**seventy typed rules peel case, tense, and pronoun endings 3. **Model lemma guess**when the stacked suffixes match no rule 4. **Generated definition**when nothing above resolves at all ::: ### Content that arrives without subtitles Most Telugu media has no subtitle file, so speech-to-text is load-bearing. Sarvam's Saaras v3 handles it, and the transcription itself is excellent — this is not where the pipeline struggles. Timestamps are. Sarvam returns a single time range per phrase, and the VTT builder splits that phrase into sentences by distributing the time proportionally across character counts. The cues are therefore estimates: the format carries milliseconds, but the values are interpolated rather than measured. Getting real precision means aligning against the audio itself, not asking the API for a finer number. ## Research ### Reading the grammar, and where the rules stop The lemmatizer was built from two grammar sources — Malik's informal description of Telugu grammar as primary, Gwynn as secondary for edge cases — read before any rules were written and generated into seventy typed suffix rules as TypeScript constants rather than string literals scattered through route handlers. The rules cover verbs: five compound classes, plus irregular అగు and a simple-past set. That is also where they stop. There is no noun path — no plural rule, nothing for case markers on nouns — and the stripper runs on every Telugu token anyway, since the only gate is whether the token contains Telugu characters. A compound like భార్యాభర్తలు (bharya + bharta + lu) matches no rule at all and falls through to the model, which guesses. The fix is not more rules. It is knowing when to strip a word and when to leave it alone — verbs yes, most nouns no — and that needs part-of-speech information the pipeline does not have. :::figure{caption="No verb rule matches a compound noun, so it falls through to the model, which guesses."} :::morphemes **భార్యాభర్తలు** భార్యా wife + భర్త husband + లు plural ::: ::: ### Dictionaries chosen for access, not authority There is nothing special about the dictionaries. Brown's 1852 Telugu-English is public domain, te.wiktionary is CC-BY-SA, both were easy to access and carried a decent corpus. That was the selection criterion. Andhrabharati and Shabdkosh were identified as upgrade paths and never pursued. If more dictionaries turn up, they go in. Ingesting them surfaced the next problem: plenty of entries that did not need to be there, suffixes among them. And a deeper one — when a learner tracks a word that is a conjugation or suffixed form of one already in the dictionary, it should not register as a new word, but it does. Two forms differing by any suffix count as two words, which inflates both the dictionary and the number of words a learner has to mark. Working out when to deduplicate those, and when context matters enough to keep them apart, is unsolved and large. Which is the honest summary of the research: even though this is the core algorithm of the tool, it is the least-researched part, because the focus went to proof of concept and workflow instead. ## Testing Forty-five test files run under vitest, weighted toward the parts where correctness is checkable in isolation — transliteration, dictionary lookup and source priority, the lemmatizer rules and stripper, subtitle and VTT handling. Those are pure functions with defined outputs, and they are where the Telugu-specific logic lives: the suffix tables, ISO 15919 with class-nasal assimilation, punctuation stripping that handles the danda. The rest cannot be tested that way. Whether a definition is right, whether a lemma is the one a speaker would actually name, whether a mined sentence is worth a card — those need judgment in Telugu, and the only person exercising it so far is the person still learning it. That cuts two ways. The grammar can be researched, so the output can be checked to a decent standard. But the vision, and the use cases that get brainstormed, are bounded by what a learner knows to look for. No Telugu speaker has used the tool yet. ## Reflection Building a tool for a language I am still learning means I can research the grammar and check output to a decent standard, but the vision, and the use cases I think to brainstorm, are bounded by what a learner knows to look for. The clearest example is a question I cannot fully answer myself: who is my market and target audience? The other thing building it has taught me is that user feedback is critical. I am an engineer at school without much business or product experience. At startup scale, the first people I would want are a marketing or business cofounder and someone who actually speaks Telugu. The core algorithm is the least-researched part of the tool, because the attention went to proof of concept and workflow instead — a large part of why it is still in development. The nearest work is a real NLP layer, an onboarding path that does not ask a learner to mark tens of thousands of words by hand, and a card and memory system of Ganula's own that syncs with Anki. Beyond Telugu, the infrastructure generalizes — mostly a matter of finding a good NLP model per language to identify word and phrase boundaries. Hindi and Tamil would keep it in the Indian-language niche; Chinese is where I would be the domain expert. That is still open. Generalizing past Telugu is the easy part. It is mostly a matter of finding a good NLP model per language to identify word and phrase boundaries, so tokens can be analyzed in the context of their source. This is a product I intend to launch. I do not know yet whether it becomes open source or a startup, but I believe in it and I am going to keep working on it. If you are working on something similar, or you are interested in this project, reach out.