AI-Generated Audio Summary
When teachers and program directors ask us why our products work the way they do, the conversation usually starts with the same misconception: that today’s AI was trained mostly on English, and that multilingual capability was added later as an afterthought. That picture is incorrect, and the difference matters for anyone teaching a second language or preparing teachers who will.
The frontier models from Meta, Google DeepMind, OpenAI, and Anthropic were built multilingual from the foundation up. Their architectures, their training data strategies, and their internal representations all assume that human language is plural. That assumption is exactly why these systems are useful for second-language education, and it is also why we have built our products at Enabling Learning, including BTLPT, TOEFL preparation, LOTE Spanish 613, our broader AI Language Labs, and Praxitude, on top of that foundation since the earliest API days.
This post is for university coordinators, program directors, and faculty who want to understand the architecture beneath the headlines, and who want to know why our model-agnostic approach is different from a generic chatbot.
The architecture is multilingual by design, not by patch
Three architectural facts shape what modern large language models can do across languages: shared embedding spaces, transfer learning, and few-shot adaptation. Each of these has direct consequences for second-language teaching.
Shared embedding spaces mean that when a model encodes the Spanish word biblioteca and the English word library, both end up represented as vectors that sit close together in the same internal geometry. The model is not translating one to the other through a dictionary lookup. It has learned that they refer to similar things, in similar contexts, with similar grammatical roles. Meta’s research line on this, from SONAR to the more recent Omnilingual SONAR, makes this concrete. Omnilingual SONAR establishes a unified semantic space spanning text, speech, code, and mathematical expressions, and on the FLORES benchmark across 200 languages it halves the cross-lingual similarity search error rate of the previous best model. On the BIBLE benchmark across 1,560 languages, it cuts the error rate by a factor of 15.
The teaching implication is direct. When a student writes a Spanish sentence and an AI evaluator reads it, the model is not running a translation pass and then evaluating English. It is reading the Spanish in a representation that already knows what the sentence is about, what tense it uses, what register it implies, and how it would land in another language. That is what makes meaningful feedback possible.
Transfer learning is the second pillar. A model that has learned grammatical structure, discourse patterns, and pragmatic markers in one language transfers a great deal of that knowledge when it processes another. This is why Meta’s Llama 4, released in April 2025, was pretrained on 200 languages with more than 100 of them represented by over a billion tokens each, and overall ten times more multilingual tokens than Llama 3. The team did not train one model per language. They trained a single model whose competence in Spanish, Portuguese, French, and dozens of others compounds rather than competes. For a university coordinator, this means an AI tutor evaluating a candidate’s Spanish writing is drawing on the same underlying linguistic competence that helps it read English, French, or Portuguese, so improvements in any one language tend to lift the others your candidates work in.
Few-shot adaptation is the third pillar, and it is the one that surprises people most. Recent work on rapid word learning, formalized in Rapid Word Learning Through Meta In-Context Learning, shows that an LLM can pick up a new word from only a handful of usage examples and use it correctly in novel contexts. The surprising part is the data efficiency: the authors trained small models from scratch on human-scale child-directed input, and these compact models matched the few-shot rare-word generalization of a much larger Llama 3 8B model that was pretrained on orders of magnitude more text. In other words, the capacity to generalize a new word from a few exposures does not require trillions of tokens. It can emerge from a training budget closer to what a child encounters before age six.
That finding is closer to how human language acquisition works than people often realize. A child does not need ten thousand exposures to biblioteca. A handful is usually enough, because the child already has a grammar, a phonology, and a world model in place. Modern LLMs have analogous internal scaffolding, which is why a few examples of a new term, a new register, or a new dialect are sometimes enough for the model to integrate it. There are caveats here, and the strongest cautionary work is the MTOB benchmark on Kalamang, a language with fewer than 200 speakers, where LLMs were asked to translate from a single field linguistics grammar book. The original results showed that LLMs could reach 44.7 chrF on Kalamang to English, compared to 51.6 chrF for a human who learned from the same materials. A 2024 follow-up showed that the LLMs were leaning heavily on the parallel sentence pairs in the book rather than the grammatical explanations themselves. The lesson is not that few-shot fails. The lesson is that the model uses what it can use, and good prompt design matters.
Speech is part of the foundation, not a separate add-on
Second-language teaching lives or dies by speech. If the model can only process text, it cannot evaluate pronunciation, fluency, or oral interaction, and it cannot help a candidate prepare for an oral exam like the BTLPT or for the speaking section of an academic English assessment.
The frontier labs have treated speech and text as parts of the same foundation. Meta’s Massively Multilingual Speech (MMS) project, documented in Scaling Speech Technology to 1,000+ Languages, supports speech-to-text and text-to-speech for 1,107 languages, identifies more than 4,000 spoken languages, and halves the word error rate of Whisper on 54 languages from the FLEURS benchmark. The project’s data strategy was creative: readings of publicly available religious texts in over 1,100 languages, providing on average 32 hours of data per language, paired with self-supervised learning. That is how you bootstrap speech for languages that have almost no commercial data. Translated to the classroom, this is why an AI evaluator can now transcribe a heritage Spanish speaker, an emergent bilingual reading aloud, or a teacher candidate practicing oral BTLPT prompts with far fewer transcription errors than a generic English-only speech recognizer would produce.
Meta’s even more recent Omnilingual MT, released in March 2026, takes the next step. It supports translation across more than 1,600 languages. The team trained two model families: OMT-LLaMA, a decoder-only specialization, and OMT-NLLB, a module inside an encoder-decoder architecture. Models in the 1B to 8B parameter range matched or exceeded a 70B baseline. They estimate the system understands more than 400 languages well enough to be useful, with non-trivial performance translating from roughly 1,600 source languages and into about 1,200 target languages. For a small language, this is the difference between having a working translation tool and having nothing.
For an educator working with emergent bilinguals, this is not abstract. It means the speech and translation capability your students need is not gated by whether their language is commercially profitable. It is increasingly accessible because the foundation models were built that way.
Where each lab actually shines, and why model selection matters
Here is where the conversation about AI in education usually goes wrong. People treat “AI” as one thing. It is not. The frontier models from different labs make different tradeoffs, and the right model for a Spanish writing evaluation is not always the right model for an English oral conversation, which is not always the right model for a low-resource pedagogical scenario.
Consider the published evidence.
Anthropic’s Claude performs strongly on translation. At the Workshop on Machine Translation 2024 (WMT24), Claude 3.5 ranked first in 9 of 11 evaluated language pairs. In a 2025 blind evaluation conducted by Lokalise with professional translators as judges, 78% of Claude 3.5’s translations were rated “good,” ahead of GPT-4, DeepL, and Google Translate. On the multilingual grade-school math benchmark MGSM, Claude scores above 90% in French, Russian, Simplified Chinese, Spanish, Bengali, Thai, German, and Japanese. The Anthropic documentation describes Claude as maintaining consistent relative performance across high-resource and lower-resource languages, with the explicit caveat that performance does drop on low-resource languages. In practical terms, this is the model we reach for when a task requires careful Spanish or English writing feedback, where the difference between “good” and “almost good” translation quality shows up directly in whether a candidate’s feedback paragraph reads as accurate to a Texas faculty reviewer.
Google DeepMind’s Gemini 2.5 emphasizes very broad coverage. Per the Gemini 2.5 technical report, the model encompasses over 400 languages via pretraining, with attention to tokenization quality, post-training data quality, and what the team calls targeted capability hillclimbing for non-English performance. On the Global MMLU Lite leaderboard, Gemini 2.5 Pro reportedly scores in the high 80s, with GPT-4o competitive on the same benchmark. The model also supports native audio output across 24 languages with a unified voice, which matters for any product that needs natural-sounding speech in multiple languages without juggling separate text-to-speech vendors. For a program serving candidates whose first languages span Spanish, Portuguese, Vietnamese, Mandarin, and beyond, the practical payoff is one consistent voice and one consistent reasoning baseline across all of them, instead of a different vendor and a different voice per language.
OpenAI’s GPT-4o, per the launch documentation, sets new high watermarks on multilingual, audio, and vision capabilities while matching GPT-4 Turbo on English text and code. The non-English improvement was the headline of that release. On a recent cross-lingual benchmark covering Cantonese, Japanese, and Turkish, GPT-4o ranked first across translation, summarization, open-domain QA, and culturally grounded dialogue, with BLEU scores of 32.5 for Turkish and 29.8 for Japanese on English-to-X translation, and ROUGE-L scores in the 36 to 38 range on summarization in the languages tested. GPT-4o’s strength on real-time audio interaction, including low-latency speech-to-speech, is a separate property that matters in conversational practice. For your candidates, that translates to oral practice sessions that feel like a conversation rather than a lag-filled exchange of clips, which is the difference between a tool a student opens twice and a tool they use for hours of speaking practice before the exam.
Meta’s Llama and Omnilingual line bring the broadest open-weight coverage. Llama 4 carries 200 languages out of pretraining. The Omnilingual MT and Omnilingual SONAR systems extend usable translation and embedding to over 1,600 and 1,560 languages respectively. For institutions that need to deploy on their own infrastructure, fine-tune for a specific dialect, or work with languages no commercial API supports, the open-weight path matters.
None of this implies a single winner. It implies that picking the model is part of the engineering work. The right answer depends on the task, the language pair, the latency budget, and the cost profile.
What model-agnostic actually looks like in our products
This is where Enabling Learning differs from generic chatbots. We are not promising “AI in education” as a slogan. We are running a portfolio of specialized products on top of carefully chosen models, with the model behind each task selected to match the task.
For BTLPT preparation, the question is how to evaluate Spanish writing, listening, and speaking against the Texas state framework with feedback that is actionable for an emergent bilingual candidate. The choice depends on whether we are scoring a written response, evaluating an oral interaction, or generating a simulated examiner conversation. Different stages, different models. The same logic applies in the LOTE Spanish 613 path, which shares some scaffolding with BTLPT but targets different domains and different learner profiles.
For TOEFL preparation, the task is academic English in writing and speaking, often for international candidates whose first language is not English and whose audio profile is not what an off-the-shelf English speech recognizer expects. We learned that lesson the hard way. The cheaper transcription models hallucinate when they encounter L2 speakers with accents. The audio coming back as text is wrong, the evaluator is then evaluating the wrong words, and the candidate sees feedback that does not match what they actually said. We standardized on the strongest available transcription model across all our products in April 2026, and the policy applies to BTLPT, TOEFL, LOTE, and Praxitude. Accuracy first.
For Praxitude, the platform is doing dual-axis evaluation: professional competence in a domain, and language proficiency in the target language, simultaneously. That kind of task benefits from models that hold long context cleanly, follow complex rubrics consistently, and can be steered with structured prompts that separate competence judgments from language judgments. Different parts of the system call different models accordingly.
For our broader AI Language Labs, the configuration changes depending on the task. A model that is excellent at Spanish writing evaluation may not be the same as the one we use for English oral conversation simulation, which may not be the same as the one we use for low-resource pedagogical scenarios that require flexible adaptation.
The point is that a serious EdTech product is not “ChatGPT with a prompt.” It is a system in which model selection is part of the design, and where the team has done the evaluation work to know which model handles which task best. That is what we mean when we say we are model-agnostic. We have no allegiance to any single lab. We pick what works for the learner.
Why this matters for university programs and faculty
If you are responsible for a teacher preparation program, an ESL initiative, or a bilingual education department, the practical implications of all this are concrete.
First, your candidates need exam preparation that meets the language at the level of the language. A monolingual English tool retrofitted with a translation layer will not evaluate Spanish writing the way a proficient Spanish-aware system will. This is not about marketing. It is about whether the feedback your candidates receive is correct.
Second, your program needs continuity across exams and across courses. If a candidate is preparing for BTLPT this semester and LOTE next semester, the cognitive load of switching tools is real. Our portfolio is designed so the same scenario logic, the same evaluation philosophy, and the same dashboard structure carry across BTLPT, LOTE, ESL Supplemental, TOEFL preparation, and Praxitude. That is why we build everything as a connected platform rather than disconnected products.
Third, your faculty deserves a tool that is honest about what it can and cannot do. Modern LLMs have weaknesses. They can hallucinate. They can sound confident while being wrong. They can underperform on low-resource languages. A serious educational platform does not pretend these limitations away. It builds in expert review, deliberate evaluation rubrics, and transparency about what the AI is doing. That is part of why we chose the term Praxitude for our platform name and Praxis Notes for the feedback layer. The framing is anti-hallucination on purpose.
Where this is heading
The trajectory of multilingual AI over the next two years is not speculation. It is visible in the public research roadmaps of every major lab.
Meta has signaled, through the Omnilingual MT release notes and the parallel Omnilingual SONAR work, that the goal is universal coverage. The team explicitly frames the work as bridging text and speech across thousands of languages, with the eventual target being any human language a learner might speak, including the low-resource and endangered ones. The strong showing of 1B to 8B parameter models against a 70B baseline points to deployment economics that put serious multilingual capability on commodity hardware within the next year or two. For an institution, that means the cost-per-evaluation curve is bending downward fast, which keeps per-student pricing predictable for programs that adopt these tools at scale.
Google DeepMind’s Gemini 2.5 technical report describes a holistic strategy in which multilingual quality is an explicit hill the team climbs every release: tokenization, data quality, modeling improvements, and capability targeting. The march from 100 to 200 to 400 languages over consecutive releases is a clear directional signal. The unified audio output across 24 languages is another concrete data point about where natural-sounding multilingual speech is going.
OpenAI has continued to push real-time audio interaction in non-English languages. The GPT-4o launch was the inflection point, and subsequent updates have lowered latency and expanded language coverage in ways that change what conversational practice can feel like. For language learners, real-time, low-latency speech interaction is the missing piece that separates “answering a chatbot” from “having a conversation.”
Anthropic’s Claude documentation and the Claude Opus 4.5 release emphasize translation quality and consistent multilingual reasoning, with explicit acknowledgment that low-resource languages remain a frontier. The next benchmarks from this lab are likely to focus on long-context coherence in multilingual settings, which matters for any system that asks a model to read a long student response, hold context across a multi-turn evaluation, and produce structured feedback.
The academic side is moving in parallel. The Rapid Word Learning work points toward models that can adapt quickly to a learner’s idiolect, a dialect, or a domain register from a small number of examples, which is exactly the kind of adaptation a tutoring system needs. Cross-lingual transfer research, such as the recent surveys on low-resource NLP, continues to refine how knowledge moves between languages. The MTOB Kalamang work and its follow-ups remind the field that prompt design and data layout still drive results, regardless of which model sits underneath.
What this means for second-language education is that the gap between “what frontier AI can do” and “what is available in classroom-grade products” is narrowing every quarter. The institutions that benefit will be the ones that have already built infrastructure to absorb new models without rewriting their evaluation logic. That is the work we have been doing since the earliest API days.
Where to start
If you are evaluating BTLPT preparation for a cohort this fall, the BTLPT product page is the right starting point. If you are looking at the broader bilingual certification picture, our TExES 165 transition roadmap walks through the changes you will face. If you are weighing an institutional commitment across multiple exams, the All-Access Semester Pass bundles the products into one decision. If your concern is ESL practice and academic English support, the ELPS Online Helper and TOEFL preparation paths are the entry points. For programs interested in scenario-based practice that evaluates competence and language together, Praxitude is where to look.
We are happy to walk faculty and program directors through any of this directly. The architecture under the hood is real, the model choices are deliberate, and the outcomes are measured. That is the foundation. Everything we build sits on it.
Enabling Learning LLC builds AI-supported preparation systems for bilingual education, ESL, and academic language assessment. Our products evaluate the right model for each task, with educator-led design and expert review at every stage.
Leave a Comment