Tag: language learning

Hindi Two-Hour Fix : I Faced Down Ergativity!

Leave a comment

You may know that I have a long-running project creating verb drill apps in lots of languages. Well, I’ve been putting off adding Hindi to VerbBlitz for years (honestly!). Not for anything you might consider ‘the hard bit’ – the transliteration of Devanagari, for example (actually really straightforward). It was because of one specific quirk that I’d built up in my head into a real hurdle: the ergativity switch in the perfective aspect.

It’s horrible linguistic jargon, ergativity, and it takes a bit of mental gym to get your head round it if you learnt about ‘subjects’ and ‘objects’ in your native language(s) in school. In a nominative-accusative (subject-object) language, like English for example, the subject gets treated the same way whether the verb is intransitive (“she left”) or transitive (“she ate the bread”). The object is the odd one out, marked or positioned differently.

An ergative-absolutive system splits things up differently. The subject of an intransitive verb – one that doesn’t take an object – is treated like the object of a transitive one – it’s ‘absolved’ of action, so it’s termed the absolutive. Meanwhile, a transitive subject gets a special marker all its own – the ergative case. If you’re wondering, that jargony label comes from Greek for worker – as it’s all about who’s doing the ‘work’ of the verb.

The Pattern in Hindi

Hindi is an Indo-European language like English, and like English, patterns like an accusative-nominative language almost everywhere. The language only switches into an ergative system in one place: transitive verbs, perfective aspect. Everywhere else, it runs on the ordinary subject-agreeing pattern, which is exactly what makes the switch feel so jarring the first time you meet it.

Take the verb khaanaa, “to eat.” In the ordinary habitual present, it behaves like you’d expect from basically any language: the verb agrees with the subject.

main roti khaataa huun — “I eat roti”
(khaataa is masculine, agreeing with main, “I”)

Now switch to the perfective — simple past, “I ate”:

maine roti khaayi — “I ate roti”

(khaayi is feminine, agreeing with roti)

Two things happen at once. First, the subject main takes an extra marker and becomes maine, so being marked as the ergative ‘worker’. Second – and this is the bit that trips people up – the verbal participle khaayi is no longer agreeing with the subject at all. It’s agreeing with roti, the object, which is feminine. Swap in a masculine object and the participle changes with it, not with you.

That’s the whole switch. Transitive verb, perfective aspect, subject gets marked with ne, and the verb hands its agreement over to the object instead.

Why it’s not actually scary

Once you see it named and isolated like that, it stops being a monster under the bed. It’s genuinely just: this one aspect, with transitive verbs, agreement moves from subject to object. Everything else about Hindi verb conjugation carries on exactly as you’d expect as a card-carrying, stock Indo-European language.

There’s also a nice way to hold it in your head if you’ve studied any Indo-European languages with a passive voice: this construction is generally thought to have started life as a passive (“by me, roti was eaten”) that then got reanalysed as an active perfective over time. Once the marker ne clicked into place as “the person who did it,” the participle just never stopped agreeing with the thing that was, historically, its grammatical subject – the object. You don’t need the history to use the pattern, but it makes the logic click.

The lesson

None of this was brain-achingly hard to learn. What was hard was the story I’d told myself about it – that it was some sprawling, thorny irregularity that would derail the whole data model. It wasn’t. It was one clean rule, confined to one aspect, with one consistent trigger.

That’s usually how it goes with the “scary” bits of a language. The size of the thing in your head and the size of the thing on the page are very rarely the same. Give it the two hours it actually needs, and more often than not it turns out to be a paragraph, not a chapter.

The MFL AI Toolkit Just Got Better: New Models for Worksheet Authoring [August 2026]

Leave a comment

Local AI is getting better. The past few weeks saw the release of several compact but powerful LLMs that make a great fit for foreign language content authoring (think worksheets and reading texts, with an even lower overhead!).

I’m using LM Studio to run these for content projects – you’ll find them all as easy downloads in the models tab. Make sure you have a bit of disk space and a fast connection – they range from 8 to nearly 20GB in size – then download and start prompting!

  • Qwen3.8-27B — Alibaba / Qwen
    Released 13–14 August 2026 · Apache 2.0 · ~27.8B parameters

    The freshest of the bunch, and arguably the pick of the litter for MFL content work. It’s the direct successor to the community-favourite Qwen3.6-27B, but now natively multimodal (text, images and video), with a 262K-token context window that stretches to 1M — plenty of headroom for feeding it whole set texts or past papers and asking it to generate comprehension exercises around them. Qwen also ships adjustable “reasoning effort,” handy for dialling compute up for tricky grammar explanations and down for quick vocab drills.

  • Muse Glimmer — Meta
    Released 10 August 2026 · Apache 2.0 · 30B parameters

    Meta’s return to genuinely open weights, distilled down from their closed flagship Muse Spark 1.2. It’s pitched squarely at agentic, on-device workflows — think batch-generating worksheet sets, checking your own answer keys, or wiring up a local pipeline that self-corrects its output — and runs comfortably on a single consumer GPU or a beefy Mac. Covers 100+ languages, which is more than most of us need, but never hurts!

  • Bonsai 27B — PrismML
    Released 14 July 2026 · Apache 2.0 · 27.8B parameters (1-bit / ternary)

    Now this is at the extreme end of “small”, and it’s attempting something quite novel. It’s an ultra-low-bit compression of Qwen3.6-27B — down to 3.9GB in its 1-bit form — small enough to run on a many portable devices (although I’ve not tried it on PocketPal yet), and comfortably on a laptop with room to spare. Quality takes a modest hit versus full-precision Qwen3.6-27B (PrismML’s own figures put it around 90% retention at 1-bit, better at the ~7GB ternary setting), so it’s more “worksheets on the train” than “your main authoring model,” but as a proof that 27B-class output now fits in your pocket, it’s a fun one to have installed.

A slightly older model that’s worth a mention alongside these is Qwen3.6-35B-A3B, a very efficient mixture-of-experts model that nominally has 35B parameters but only about 3B are active at any one time via a clever routing trick. It’s noticeably faster than Qwen3.8-27B for similar output quality in my own testing, and it’s been my workhorse for some time. It’s worth keeping an eye out for a future MoE version of 3.8 built the same way — that could be genuinely excellent.

Why Small Models?

Using smaller models locally addresses many of the environmental and privacy concerns around AI use. All the prompt compute happens privately on your local machine — no student data, no draft worksheets, no half-finished reading texts ever leave your laptop. And despite the tiny footprint, they’re still excellent at producing native-quality text across dozens of languages, which was frankly unthinkable on consumer hardware even a year ago.

None of these is going to replace a careful human eye on the final worksheet — but as first drafts, distractor generation, or a quick “give me five more sentences at this CEFR level,” they’re genuinely useful additions to the toolkit. Worth an afternoon of poking around, at the very least.

Happy Commonwealth Games 2026 – Enjoy The Languages, Glasgow!

Leave a comment

A busy one this week, as many of us in Scotland (and beyond) are preparing for the Commonwealth Games in Glasgow, either as volunteers or visitors. The 2026 edition of the games will be filling the city with sporting joy from 23rd July to 2nd August.

And there’s a language connection there, of course – I couldn’t hide my enthusiasm at that very fact when it was Birmingham’s turn to host the Games in 2022. A huge number of languages are spoken across the Commonwealth, with some countries hanging on to a staggering level of linguistic diversity. Nearly a thousand languages are spoken in Papua New Guinea. Nigeria has over 500. India with over 400, Australia breaking 200, and Cameroon at nearly 300 round off the medal table.

Some of those, of course, will be represented by athletes and spectators coming to Glasgow. It’s a lovely thought to think that the event might bring Arrernte, Hiri Motu or Fulfulde to the streets of the city.

As for me, I’m looking forward to a couple especially. There’s Greek, obviously, thanks to Cyprus. We should hear a fair bit of Swahili about, with visitors from Kenya, Rwanda, Tanzania and Uganda. And francophone countries are in the mix too, with Canada, Cameroon, Mauritius, Vanuatu, and of course, Rwanda again. I’m a little giddy at the thought of hearing Maltese, too – not one I’ve ever tried to learn, but I’ve had a vintage copy of Teach Yourself Maltese on my shelf for years that wants to persuade me otherwise!

And as well as all the Scots, of course, there’s a wee treat for Gaelic speakers and learners – BBC Alba are covering the games in the language.

If you’re in Glasgow, have fun listening out for those languages. And wherever you’re following the fun, have a wonderful Commonwealth Games 2026. Find out more at the official site!

Snarky Marketing vs. Real Data: Duolingo Really Works!

Leave a comment

It’s become almost trendy to knock Duolingo in the last couple of years. In fact, positioning a new language learning product as the “anti-Duolingo” has become a fairly standard marketing tactic.

And while there are always reasons to debate a language learning method – no one approach is perfect in every way for every person – it’s good to know the difference between a social media pile-on and legitimate criticism.

A new paper has provided some more objective balance on that.

Published in Language Learning & Technology, a study by Smith, Jiang, and Peters (2024) took a rigorous look at how effective the app actually is for independent learners. Instead of just relying on anecdotal feelings, they measured outcomes across a comprehensive range of linguistic abilities. We’re talking both receptive skills (reading and listening) and productive skills (writing and speaking), alongside vocabulary, grammar, and pronunciation.

They tracked 48 independent learners tackling the Spanish course over a three-month period. The findings? After putting in an average of about 27 hours of study, participants showed significant improvement across all ability measures. A small sample, for sure, but 27 hours is remarkably low for meaningful gains – less than 15 minutes a day. For an app coming under flak for meaningless gamification, that is an excellent piece of defence.

When gamification earns its keep

Working in educational software development myself, I know how tricky it is to balance engaging user interfaces with genuine pedagogical value. It’s one thing to build a platform that people want to log into every day, but it’s another entirely to prove that it’s actually teaching them something substantive. This paper validates that the gamified approach, when engaged with consistently, really does translate into measurable linguistic gains. The researchers specifically noted that session completion, accuracy rate, and an overall positive user experience were directly linked to this observed growth.

This challenges one of the most common criticisms you hear: that apps like this only teach you how to pass their specific tests, or only build passive recognition. The fact that productive abilities like speaking and writing improved significantly is a huge win, as these are notoriously difficult to scaffold outside of a classroom or immersion environment.

Of course, this isn’t to say any app is a magic bullet. Twenty-seven hours of app time won’t instantly make you fluent enough to debate current affairs over a coffee. I’m a Duolingo user, and get frustration and satisfaction from it in, sometimes, equal amounts. It never hurts to spread your screen time over other language learning apps, either.

But as a tool for building a solid foundation and keeping learners consistently engaged with the target language, this fresh evidence is robust. It turns out that beneath the relentless notifications and leaderboard leagues, there is a some sound pedagogy at work.

The next time a social media ad pops up claiming they’ve “fixed everything the green owl gets wrong”, it might be worth taking it with a pinch of salt. Any tool that gets people to consistently show up, put the hours in, and tangibly improve is something worth celebrating!

The Perfection Trap: Why ‘Sounding Native’ Isn’t the Ultimate Accent Goal

Leave a comment

A paper appeared this month in Language Teaching that reminds us all – in the classroom and learning at home – that accent isn’t everything.

Is English pronunciation teaching in China ready for the Intelligibility Principle? (Xue & Wu, 2026) looks specifically at English language teaching in China, and the pressure that aspirations of nativeness can put on students. For a polyglot community focused on becoming the languages we learn, this seems very familiar. The pursuit of a flawless, native-like accent can often feel like the ultimate prize – anything less is an imperfect near miss. Whether you are tackling the phonology of German, Spanish, or diving into a new challenge like Bulgarian, it is easy to obsess over perfectly mirroring native audio.

However, the authors gently remind us that this pursuit can be entirely unnecessary.

Historically, language instruction has been dominated by this Nativeness Principle, which dictates that learners should attempt to sound as close to native speakers as possible. It’s a hangover from a quite Victorian perfectionism in learning – learn by rote, never forget.  Yet, absolute native-speaker norms are an unrealistic objective for many learners – especially older ones (as much as it stings to hear that pushing 50!).

Instead, a more practical framework is the Intelligibility Principle, which argues that the primary pedagogical goal should simply be mutual understanding.

Does This Matter for the Polyglot Community?

Absolutely. The pressure to achieve a flawless accent can actively hinder our progress. An overemphasis on native-speaker models can add to language anxiety and can decrease a learner’s willingness to communicate. Not surprisingly, many of us have an aversion to feeling a fool – it’s great to push back against that as a language learner, but it’s not always easy.

The reassuring truth is that speech can carry a heavy foreign accent and still remain highly intelligible to listeners, and therefore functional. A lot has been made of Zelenskyy’s excellent communicative skills despite a strong accent, for example. Furthermore, because English and many other languages act as global lingua francas, learners frequently communicate with other non-native speakers. Therefore, a native accent is neither necessary nor particularly beneficial for mutual understanding.

The goal isn’t to erase our linguistic identity, but to ensure our message is clearly received.

Practical Steps: What Should Independent Learners Focus On?

The paper is quite rich on practical suggestions in this regard. When you are building your corpus of vocabulary or grinding through daily flashcards, it helps to know exactly which phonetic features actually impact comprehension. Instead of trying to master every subtle nuance, prioritise the elements that carry the highest functional load:

  • Vowel accuracy: Both the length (quantity) of a vowel and the quality of the vowel play significant roles in whether a word is understood.
  • Consonant clarity: Avoiding the simplification of diphthongs and preventing the conflation of voiced and voiceless plosives are critical steps for maintaining clarity.
  • The bigger picture: Suprasegmental features, which include stress, rhythm, and intonation, have a remarkably strong influence on how a listener perceives an individual’s oral proficiency.

Tools and Tactics for the Daily Grind

As independent learners and language hobbyists, we have to direct our own study routines. Here is how we can implement an intelligibility-first approach:

  • Embrace accommodation strategies: In real-world conversations, mutual understanding is often achieved through spontaneous adjustments like simplification and paraphrasing.
  • Leverage technology: Computer-assisted pronunciation training and artificial intelligence-powered tools offer instant, data-driven feedback and personalised suggestions for self-monitoring.
  • Recognise algorithmic bias: If you are using—or even developing—language learning apps, it is important to remember that commercial automatic speech recognition systems frequently exhibit bias, showing significantly higher error rates for non-native accents.

Ultimately, dropping the demand for native-like perfection takes a massive weight off the language learner’s shoulders. By focusing on practical intelligibility, clear suprasegmentals, and flexible communication strategies, we can speak our target languages with far more confidence and far less anxiety.

Have you evolved from learner perfectionism to more intelligibility-based goals yourself? Was it hard to let go, or did you feel a sense of relief at letting go? Let us know in the comments!

A mic stands on a stage, ready for #EdFringe

#EdFringe for Language Learners, 2026 Edition

Leave a comment

#EdFringe is here again, and true to form, there’s something for everyone. Language learners and European culture vultures are no exception, of course, with some proper treats in this year’s rich programme.

As per last year (and the year before, and the year before that… I’ve been keeping tabs for a while now), there’s a good balance between shows in the target language as well as shows about target language counties. Here’s my watch list – but be sure to have a browse too, and let me know if I’ve missed any must-sees!

French

Music

Le Vent du Nord – Québecois progressive folk from a well-regarded five-piece.

Mary, Queen of Scots – Queen of 3 Kingdoms (Marianne Beate Kielland & Ben-San Lau) – French arias mingle with Scots and English in a programme celebrating the life of Scotland’s famous queen.

Afternoon Arias (Brian Bannatyne-Scott, Beth Taylor and friends) – French classical highlights from Berlioz to Debussy.

La Chatte Chanteuse (Kat Brooks) – It honestly wouldn’t be #EdFringe without a bit of Piaf. There’s that and more in this hour of chanson from Kat Brooks.

Theatre

Madame La Mort (Full Moon Theatre / Labyrinth Productions) – a radical reimagining of Rachilde’s French Symbolist play (in translation)

Comedy

Tori Morancay – Le French C’est Freak – An anglophone set that nonetheless plunges straight into the francophone world!

German

Music

Handel – Nine German Arias (Angela Hicks)  – Soprano Angela Hicks and ensemble present some of Handel’s few German-language works.

JS Bach (Aidan Jones) – Pianist Aidan Jones plays and presents, taking the audience on a winsome tour through the life of the great composer.

Scottish Lieder (Brian Bannatyne-Scott) – A lovely crossover presenting music of Schubert, Schumann, Loewe and Strauss, inspired by Scottish poets.

Comedy

Michael Brunström: William Tell vs the Algorithm – Swiss surrealist comedy (in English) that takes aim squarely at the Swiss background of this award-nominated return performer.

Jürgen Strack: Achtung! The Only Sauerkraut in Town – Riffing on his Germanness (with shades of Henning Wehn here), Strack has won fans for his sheer originality.

Spanish

Music & Dance

Sobremesa – Where Words and Music Meet at the Table (Nus Duo) – Billed as an interactive musical experience where audience and artists shape something wonderful from Spanish and Latin American texts.

Sounds of St Cecilia’s III: Spanish Flavours: Dance, Fire and Elegance (Cokus Duo) – A harpsichord-led exploration of the music of 18th-century Europe, with a nod to both France and Spain.

Flamenkids (TuFlamenco) – Flamenco is as permanent a fixture on Spanish #EdFringe as Piaf for the French. This family-friendly show introduces the rhythms of Spain in an hour-long show that sounds wonderfully interactive.

Alegria Flamenca (Alba Flamenca) – Appetite freshly whetted, if you’re now hooked on the flamenco check out this vibrant show, with a nearby bar on hand for tapas and drinks!

Theatre

Bull / Fight – Fresh Edinburgh ensemble Mythography presents this odyssey through Lorca’s Spain (in English).

Comedy

Escocia con Ñ (Jotace Loaiza) – Scottish life retold through Spanish eyes, in Spanish! An excellent (and rare) opportunity to attend a full Spanish-language set during the festival.

Mi Casa Es Su Casa (El Purnell) – Billed as a true ‘duo-lingo’ act (hope that’s been run past the owl!), laughs are promised for hispanists and non-hispanists alike.

And the rest…

With hundreds and hundreds of shows, I can’t possibly do the whole programme justice in one short blog post. While I’ve focused on French, German and Spanish, there’s plenty else there too, from the very tempting Four Courses of Italian Song (Anna Vanosi) and the intriguing I Can Make You Italian in 55 Minutes (Stefania Licari), to the Scottish and Norwegian sacred music presentation The Maid of Norway (Nordic Voices Norway) and online viral comedy hitster Thor Stenhaug. I’ll certainly be trying to tick some of those off my list this August, too.

Is there something you’ve bookmarked to see without fail? What other languages are represented in the listings? Let us know in the comments!

Forever the Optimist : Hugo’s In Three Months Language Courses

Leave a comment

I loved Hugo’s In Three Months books as a language-obsessed teenager. Slim volumes, tiny, dense chapters… And of course, that promise that you’d reach some level of fluency in a quarter of a year.

I’ve collected a fair few of them over the years. My first was Italian in Three Months, of Hugo’s late-80s dark blue incarnation. I bought it ahead of a school wind band trip to Venice, and probably made my way through about a third of it before the trip was over, I packed my tuba away, and faddishly drifted onto the next project.

They’ve had a much longer history than that. Hugo’s Language Learning Books pops up in 1950s UK, and quickly starts churning out titles for speedy learners. Like other language series in the mid-20th Century, they expanded their Simplified System rapidly across languages, and became bookshop staples.

Expansion, incorporation and multiple reincarnations

By the end of the 80s, they’d already moved well beyond French, German and Spanish. The second-hand trail on eBay shows that spread, with the publication of courses from Arabic to Scottish Gaelic (still a solid reference for Gaelic grammar if you can get a copy). Curiously, the first foray into Japanese dropped the ‘in three months’, instead going for Japanese Simplified – the confidence wobble didn’t last long, as Japanese in Three Months is the title that made it into the 90s!

The In Three Months books have never really gone away. The series was picked up by Dorling Kindersley (DK), gaining a splash of the prototypical DK colour and gloss. While the range of languages is a little shrunken now, they’re still going strong, now as part of Penguin’s catalogue. Gone are the boxes of cassettes, replaced by online audio. But the familiar format remains: tight, reference-style chapters giving that “all the essentials for very busy people” vibe.

It’s in that spirit I picked up one of the new editions recently – appropriately, Italian in Three Months again, this time in its fancy new green sleeve. It’s a refresher ahead of a trip to Milan for a conference, and the perfect choice for that – not too heavy (for the suitcase or the reading).

But coming full circle like this takes me right back, and I can sense the excitement I felt in that titular promise all those years ago. Somehow, that promise still works. Long live the In Three Months series – may they continue to lure language nerds to their next obsession!

Gramophone Language Courses: The Original Multimedia Learning

Leave a comment

If you’ve ever wondered about the origins of the multimedia language course, then some newly published archive material might surprise and intrigue you. The British Newspaper Archive recently added the early 20th-century title Sound Wave magazine to its growing catalogue. This record review title served phonographic fans from 1907 to 1933, and it’s surprisingly full of language learning history.

In those days, of course, it was the gramophone that reigned supreme. Recordings on the new flat disc format had been around since the late 1880s, but by the first decade of the 20th century, gramophones had become affordable enough for middle-class households. Sound Wave dates from that early tech spread, the publishers no doubt spotting a gap in the market for listening recommendations.

Only it wasn’t just music. What we’d now recognise as audiobooks was already in circulation – outfits like The Talking-Book Corporation were pumping out gramophone literature for adults and kids. There were elocution resources for improving one’s spoken English, too. One particularly enticing release was this special set of discs with the voice of Bernard Shaw himself (life imitating art – his own art!).

And language learning was there from the start too, in the form of regular ads from the Linguaphone Institute.

A 1907 advert from the Linguaphone Institute in the magazine Sound Wave

A 1907 advert from the Linguaphone Institute in the magazine Sound Wave

Linguaphone – a brand built on gramophone

Linguaphone is a real heritage brand for language learners, and pops up all over the newspaper archives. It started up in 1901, and is still going today – you may have come across their language training centres. Competition may have widened since then, but for over half a century they were the first word in audio course materials.

Testimonials in this 1927 edition vouch for their success. One C.B. of London reports that the Spanish course made travels “much easier and cheaper than they would otherwise have been”. A reviewer in 1929 praised the “French as it is really spoken” in a dialogue set in a hairdresser’s, on record no. 21 of that set. Yes, record 21 – these sets could run into dozens of discs, and usually shipped in a hefty, solid case.

No wonder they came with an equally solid price tag. In 1907, a box would set you back £3 and three shillings, easily several hundred pounds in today’s money. You can still pick them up second-hand today, and for much less – a lovely bit of language learning history.

Proto Language Lab

Beyond the well-heeled turntable owner, the gramophonic method wasn’t just for individuals; it was used in classrooms too. In 1914, a Leicester teacher, Mr. Cunfliffe, introduced records into his lessons at the Working Men’s College, to great success. One particularly modern-seeming innovation of Mr. Cunliffe’s was the provision of “24 pairs of hearing tubes” for the students! In this way, one element of language teaching that seems so late 20th-century, so proto language lab, had its roots decades before tape reels and CDs.

The BNA‘s inclusion of Sound Wave offers some lovely insights into the history of language learning and teaching. There’s doubtless much more to find in there. Let me know in the comments if you come across any other gems!

Screenscot of Cell to Singularity, an immersive casual clicker game available on Steam.

Cell to Singularity : Casual Play for TL Immersion

Leave a comment

Osmosis isn’t just for cells – it’s for language learners too! Soaking up target language simply by placing it in your everyday line of sight is one of the most effective strategies for fluency. From your instagram feed to cosy telly-watching, consolidation can be about throwing more of the things you love in your way.

Gaming is another entz stream that is really easy to target language-ify, since many titles have multiple language options. The Steam platform is a particular goldmine here – a huge multi-platform marketplace, with loads of free-to-play offerings. The trick is to find quite text-heavy games with dialogue and interactions, exposing you to as much content as possible in-play. There’s honestly something for everyone here, from word games to fully-fledged RPG.

This week, I chanced across a casual clicker on Steam that has been working its quiet way into the hearts of users since its inception in 2018. It’s Cell to Singularity, a game that simulates the blossoming of life on Earth, from eukaryotes, to jellyfish, to humans (and beyond). It’s the kind of game you can have running inconspicuously in the background while you work, slowly developing and growing like a bonsai that needs occasional tending. Very Zen.

Screenscot of Cell to Singularity, an immersive casual clicker game available on Steam.

As you can see from the screenshot, it’s also a great way to revise the building blocks of life. That’s the root educational application the game has been feted for, covering evolutionary biology in a fun, laddered way. Switching my interface to German gives me a ton of fun natural world vocab.

Beyond word level

But the game is also full of conversational exchanges you have with the ‘supercomputer’ running your life simulation, as well as Wikipedia-style descriptions of all your finds. In short, it supports word, sentence and text-level language skills in a rich, engaging environment. What more could you ask for?

Screenshot from Cell to Singularity showing dinosaurs

The range of languages available right now is already impressive. Not only the ‘mainstream’ school ones, but also Korean, Japanese, Polish and Portuguese, amongst others.

Screenscot of the language options in Cell to Singularity, an immersive casual clicker game available on Steam.

Cell to Singularity currently has an 89% positive rating from thousands of Steam users. I wonder how many of them are playing to improve their target language? Hopefully I’ve enticed a few more of you to do just that!

Speaking dialect in the mountains

AI and Dialect: Can You Nudge a Language Model Off the Standard Path?

Leave a comment

One of my quiet worries about the AI revolution in language is standardisation. When a model trained on billions of documents produces text, it tends to sand everything down to some kind of weighted average – grammatically impeccable, register-appropriate, but utterly devoid of local colour. And where’s the fun in that? You know my thoughts on dialect and language learning.

But here’s the thing: despite the odds, the standard grey slop isn’t a mathematical inevitability. Large Language Models are trained on billions of words. And that material will include plenty of non-standard language, too, simply by dint of the extraordinarily broad sweep of the training brush. Dialect literature, forum posts, subtitles, transcripts – there’s room for non-standard, more colloquial language in every context of the model’s background.

Which raises an interesting question: can we nudge AI output away from the standard, and towards something more authentically regional?

I’ve been playing around with exactly this, and the short answer is: yes – and it’s surprisingly decent at it.

Stepping off the well-trodden path

I’ve said from the start that AI is built for language learning content creation. There aren’t many fields where content accuracy doesn’t matter too much – as long as the language is correct and authentic-sounding. One of my go-to moves is pushing the LLM for a more colloquial register. Standard written language is fine for reading comprehension, but if you’re teaching people how to actually speak, you want contractions, ellipsis, informal vocabulary – the stuff that sounds like a real person. AI handles this pretty well once you prompt for it explicitly.

Create a model German dialogue on the topic “Reality TV”. The language should be contemporary, colloquial and natural-sounding German.

Claude did a pretty good job of this:

  • Lena: Ey, hast du letzte Woche „Love Island” geschaut? Ich konnte gar nicht aufhören – ich war bis halb zwei wach.
  • Kai: Natürlich hab ich das geschaut – wär ich ja bescheuert, das zu verpassen. Diese Tiara und Marco… ich meine, komm schon. Die spielen da nur rum.
  • Lena: Genau! Marco tut total so, als wäre er an ihr interessiert, aber das ist doch alles für die Kamera. Ich find das so zum Fremdschämen.
  • Kai: Ja, aber genau deswegen schaust du’s doch! Gib’s zu. Du liebst den Drama.
  • Lena: Okay, okay, ich geb’s zu. Aber ich find’s trotzdem fragwürdig, wie die Leute da so bloßgestellt werden. Manche wirken echt überfordert.
  • Kai: Absolut. Ich glaub, vielen ist nicht bewusst, wie sehr die Produktion das alles lenkt. Die schneiden das so zusammen, dass du denkst, die Person ist voll die Zicke – aber vielleicht war das gar nicht so gemeint.
  • Lena: Stimmt schon. Und dann wird die Person im Netz fertiggemacht. Ich hab mal gelesen, dass einige danach richtig psychische Probleme hatten.
  • Kai: Das ist so traurig. Und trotzdem melden sich jedes Jahr tausende Leute an. Ich versteh’s irgendwie – du willst berühmt werden, schnelles Geld, Follower… Der Lockruf ist halt stark.
  • Lena: Stimmt. Und seien wir ehrlich – wir beide gucken’s ja auch. Also sind wir Teil des Problems, oder?
  • Kai: Pff… ich würd sagen: wir sind Teil des Unterhaltsystems. Klingt besser. Jetzt mach die nächste Folge an.

Some great non-standard, everyday features in there. Colloquial contractions like -e dropping on verbs, decent use of particles like halt – not a bad job at all. Claude even threw in a bunch of usage notes without asking, too.

But colloquial is still, broadly, floating just beneath a standard umbrella. What if we go further?

Adding Local Flavour

When I’m probing the edges of what AI can do, I like to start with varieties I can actually judge. Austria is my German-speaking region of choice in this regard, being my year abroad destination of old – I still have a huge soft spot for Kärnterisch. So here’s the tweak:

Create a model German dialogue on the topic “Reality TV” between two young Austrian speakers from Kärnten. The language should be contemporary, colloquial and natural-sounding Austrian German.

Here’s what came back:

  • Magdi: Oida, hast du gestern „Love Island” g’schaut? I kann nimma! Die Lea hat den Stefan komplett ausg’spielt, echt a Wahnsinn.
  • Tobi: Na freilich, i schau des jo seit Wochn! Aber geh, die Lea, die tuat doch nur so. Die weiß ganz genau, wos sie macht — des is alles g’spield.
  • Magdi: Ja eh, ober trotzdem — der Stefan is so a Trottel! Wie kann man so blöd sein, gell? Der glaubt ihr jo alles.
  • Tobi: Najo, i mein, de san jo auch so a bissi animiert zum Streiten, oder? Die Produzenten woin das jo so. Ohne Zoff läuft des ned.
  • Magdi: Des stimmt scho, ober i find des trotzdem scheißn. Die tean so, ois ob’s echt wär, und dann geht’s nur ums Follower kriegen. Danoch san alle auf Instagram mit Werbung zuagekleistert.
  • Tobi: Haha, stimmt! Drei Wochen nach der Sendung verkaufen’s alle irgendwelche Protein-Shakes. Gott sei Dank ned mia. [lacht]
  • Magdi: Ober weißt wos? I schau’s trotzdem weiter. I bin schon voll drin, i muss wissen, ob der Kevin die Kathi kriegt oder ned.
  • Tobi: Na sag i dir: die kriegen sich ned. Die Kathi is viel zu schlau für ihn. Die geht da raus, macht a paar Storys, und dann war’s des.
  • Magdi: Oida, du bist so a Zyniker! Vielleicht is es jo echt?
  • Tobi: Magdi. Des is Reality TV. Des is so echt wie mei Oma beim Wrestling. [grinst]
  • Magdi: Haha, okay okay. Ober schauen tust’s trotzdem, gell?
  • Tobi: … Ja. Natürli. Um neun bin i dabei. Sag’s ned weiter.

If you’ve spent any time in Austria, this is probably just as joyful to you as it is to me. It’s added in so much southern colour, from the tag particle ‘gell’ to the reduced i from ich and past participles in g’. There’s even some fun wordplay in there (as real as my gran at wrestling!), so it’s not only dialect, but character. It’s laid it on thick!

There’s an element of pastiche to this, for sure. Nobody will claim this is indistinguishable from an authentic speaker of Klagenfurt. But it’s genuinely not bad, especially if you’re after a solid introduction to local speech features.

And then there’s Trøndersk…

That’s all very well for ‘mainstream’ foreign languages like German. To push the envelope a little more, it’s worth playing around with some lesser-learnt languages – Norwegian is always a good choice.

For my Norwegian test, I went with Trøndersk, the dialect cluster associated with the Trondheim region and much of central Norway. It’s one of the most immediately recognisable Norwegian dialects, with distinctive pitch accent patterns, characteristic -a verb endings, and a few phonological quirks that mark it out clearly from the Oslo Bokmål you’ll meet in textbooks.

How good a job would Claude do of this challenge?

  • Marte: Oi, har du sett at det e ny sesong av Farmen nå? Æ begynte å se på den i går kveld og klarte itj å slutt.
  • Jonas: Åja? Æ har itte sett mye reality-greier på ei stund, men Farmen e jo noe anna. Det e liksom ordentlig underholdning, da.
  • Marte: Ja, det e noko med den serien! Og den nye dama fra Steinkjer – ho e jo helt vill. Ho klaga på absolutt alt og alle og gjor ingenting sjæl.
  • Jonas: Haha, klassisk. Det e alltid nåkka en sånn type på realitygreier. Produsenten e nødt til å ha med dem, skjønner du? Ellers e det jo kjedelig.
  • Marte: Jaja, men det e jo litt ekkelt å innrømme at man liker det, liksom. Æ satt der og ropte på skjermen som et idiot.
  • Jonas: Haha! Ka sa du da?
  • Marte: Noe sånt som «koffor gjør’u det, din tulling!» Men mora mi satt rett ved sia og hun så bare på mæ som om æ var gal.
  • Jonas: Uff da, haha. Ka syns ho om programmet da?
  • Marte: Ho e heilt avhengig sjæl! Det e liksom heime hos oss at hele familien samles rundt skjermen på tirsdager. Det har vorre sånn siden æ va lita.
  • Jonas: Okei men ka med all den andre reality-søpla da? Eks på stranda og sånt? Det e jo bare tull, syns æ.
  • Marte: Ja nei, det gidder itj æ. Det e jo bare folk som vil bli kjendis og som itte har nåkka å by på. Farmen e annleis fordi det faktisk krev noe.
  • Jonas: Sant nok. Kameraten min i Ålesund e helt hekta på Paradise Hotel derimot. Han ser på det heile tida.
  • Marte: Seriøst? Koss klare han det? Æ holdt ut to episoder og ga opp. Alle folkan der oppe e jo helt… æ veit itte. Tomme, liksom.
  • Jonas: Haha, han sier det e «sosialt eksperiment». Æ tror bare han har lyst på drama heime fra sofaen sjæl.
  • Marte: Hahaha! Ja det e nok det. Uansett — skal du se Farmen med mæ på fredag? Vi kan bestille pizza og le av alle.
  • Jonas: Ja, det høres bra ut! Men du bestemme pizza-topping. Sist gang valde du masse squash og det va itte greit.
  • Marte: Oi, squash e jo godt da! Men okei, du får velge. Bare itte ananas.
  • Jonas: Deal. Vi ses fredag!

As far as I can tell – and I’m not so well placed here, I must admit – it’s done a solid job. The k- question words (rather than hv-), the reduced æ for jeg… Even the cultural references, selecting two actual Norwegian shows for the exchange. I’d trust this as a bit of pre-Trondheim trip prep!

That said, if you speak a Trøndersk dialect, what’s the verdict? Does this read as a reasonable attempt, or is it giving you second-hand embarrassment? I’d genuinely love to know. (Drop a comment below, or find me wherever I am on the internet this week.)

AI and Dialect for Language Learning

AI platforms clearly have some representation of dialect. Awareness is the wrong word for it – anthropomorphising LLMs and all that – but you know what I mean. The patterns are in the training data; the question is whether a prompt can reliably surface them, or whether the model’s default pull towards the standard is too strong.

The two languages I explored here seem to work well. The pull to parody is always something to watch, particularly when written material in a dialect tends to be humorous, rather than in-community. And rarer or more exclusively spoken varieties will no doubt be much more troublesome. But the more linguistic scaffolding you give the model – the more you treat prompting as a kind of sociolinguistic briefing – the better the output is likely to get.

That standard grey slop, it seems, has more texture than it first appears – you just have to know where to poke.