cs.CLSep 12, 2026
SaveLLMersion: A Local-First AI Agent Framework for Low-Cost Home Language Learning toward Educational Equity
Organizations: Dept. of Computing Sciences Texas A&M University–Corpus Christi Corpus Christi, TX, USA · Dept. of Electrical Engineering and Computer Science University of Missouri Columbia, MO, USA · Dept. of Computer Science and Engineering University of California, Riverside Riverside, CA, USA · Dept. of Mechanical and Aerospace Engineering University of South Florida Tampa, FL, USA ·
Abstract
Artificial intelligence helps education most where an essential provision has been rationed by cost. For language learners that provision is a teacher's voice, which binds listening, reading, speaking, and writing into one act. Published evidence shows why most learners lack it, from a global shortage of 44 million teachers to heavy household tutoring bills, and why technology has not substituted for it: computer-assisted language learning proved effective but narrow, applications presuppose connectivity 2.6 billion people lack, and One Laptop per Child's randomized evaluation found that hardware without capable software teaches nothing. We distill eight difficulties and four binding constraints, and argue that small open-weight models dissolve the last: a complete four-skill stack now fits a $200-class laptop and, on community measurements, generates at the pace speech is consumed, for about one US cent of electricity per study hour. We therefore propose LLMersion, a scheme for AI for education that runs entirely at home, over the learner's own documents, with an AI-written, AI-understood, AI-updated codebase anyone can customize; present LLMersion-1, a released open-source prototype (https://github.com/QM378/LLMersion ); and outline the vision of a private learning agent.
Figures & tables
| Difficulty | Why it persists | Prior answer | Its limit | This scheme |
|---|---|---|---|---|
| Absolute teacher scarcity | Demand outstrips trained supply; attrition nearly doubled 2015–2022 ( UNESCO, 2024 ) | Broadcast courses, MOOCs, apps | Presuppose connectivity 2.6B people lack ( International Telecommunication Union, 2024 ) | The teacher’s four roles as software on hardware already owned |
| Teachers’ own oral proficiency | Teachers formed under grammar/exam paradigms; majorities test below required levels ( Coleman, 2009 ) | Broadcast native audio | One-way; no feedback on the learner | Synthesized model voice over any chosen text |
| Individual speaking time rationed | Classes of 21+ students; teacher talk dominates ( OECD, 2025a ) | ASR pronunciation drills | Pre-deep-learning ASR doubled errors on L2 speech ( McCrocklin, 2026 ) | Unlimited private spoken turns with feedback |
| Method locked to written exams | High-stakes tests reward reading and grammar ( Liu and Ren, 2024 ) | Test-preparation software | Reinforces the lock | All four skills exercised over the same document |
| Reading outruns speech | Without oral input/output, print decouples from phonology ( Costello et al., 2021 ) | Audiobooks, subtitles | Fixed library; no interaction with the learner’s material | Read-aloud, then spoken discussion, of the learner’s own text |
| Proficiency tracks wealth and geography | Oral practice is bought privately; +34% wage premium in India ( Azam et al., 2013 ) ; listening tests widened China’s rural-urban gap ( Li et al., 2024 ) | Subscription applications | Recurring fees are a regressive filter | One-time commodity hardware; near-zero marginal cost |
Table 1: Eight difficulties of second-language acquisition under resource scarcity, the mechanism sustaining each, technology’s prior answer, its documented limit, and this scheme’s response. Sections 2 – 4 develop each row with sources.
| Where | Anchor (data year) | Scope note |
|---|---|---|
| Korea | 29.2T won total; English top subject at 248k won/mo (2023–24) ( Statistics Korea and Ministry of Education, Republic of Korea, 2025 ) | Official national survey |
| China | RMB 8,438/yr mean per tutoring household; 46.2% participation in first-tier cities (2019) ( Wei, 2024 ) | Nationally representative survey; sector largely closed by 2021 decree ( State Council of the People’s Republic of China, 2021 ; Lyu and Lam, 2025 ) |
| India | 11.8% of household education spending; participation 27% by 2025 ( National Statistical Office, Ministry of Statistics and Programme Implementation, Government of India, 2019 ; National Statistical Office, Ministry of Statistics and Programme Implementation, Government of India, 2025 ) | Official NSS rounds; no per-subject breakdown |
| Vietnam | 77% of upper-secondary students tutored (survey era) ( Dang, 2007 ) | Peer-reviewed household data |
| Japan | ¥61B/yr, conversation schools (2024) ( Ministry of Economy, Trade and Industry, Japan, 2024 ) | Ministry survey; top-firm sample, not full census |
Table 2: What families pay: one official or peer-reviewed anchor per country. Figures are for private supplementary education; scope notes matter and are part of the claim.
Figure 1: The scheme. Dashed nodes are replaceable open-weight models chosen for minimum hardware demand; everything runs on the learner’s machine over the learner’s own material. An instance is described in Section 6 .
| Component | Params | Size (Q4 / weights) |
|---|---|---|
| Tutor LLM | 1B / 3B | 0.81 / 2.02 GB |
| Voice (TTS) | 82M | 0.33 GB |
| Ear (ASR, “small”) | 244M | 0.47 GB |
| Bilingual (MT) | 77M | 0.30 GB |
| Whole stack (3B tutor) | 4 GB | |
| Cheapest new laptop memory | 8–16 GB |
Table 3: The complete four-skill stack against commodity memory. Sizes are published file or model-card figures ( hugging-quants (Hugging Face), 2024 ; hexgrad, 2025 ; OpenWhispr, 2026 ; Tiedemann and Thottingal, 2020 ; Walmart.com, 2026 ) ; the resident total leaves headroom on an 8 GB device. Throughput on the cheapest CPUs is community-measured (see text), not peer-reviewed.
Figure 2: The LLMersion-1 reader during a read-along session, as seen by a Chinese-speaking learner of English. Interface labels are in the learner’s first language. From left to right the toolbar reads: open file; voice (here the Kokoro voice “Heart”, General American); playback speed; reading mode (whole paragraph); Chinese translation on or off; vocabulary book, with one saved word; writing practice; conversation; and a light or dark theme switch. The English paragraph is the text being read aloud, and the yellow band marks the current reading position as the voice proceeds. The grey text beneath it is a machine translation of the same paragraph into Simplified Chinese, so the learner can check comprehension without leaving the page. The bar at the bottom holds the playback controls and the position in the document.
Figure 3: The speaking and writing modules (interface text in Chinese, the learner’s first language). (a) The conversation module before a session. The heading reads “Before you speak, choose a starting point” and the line below it, “Whichever you choose, the tutor will deliberately use words from your vocabulary book, and every sentence you say will get pronunciation feedback and corrections along the way.” The three options are free talk, a chosen topic, and talk about this document, followed by a start button. The strip labelled “your words” holds endowed , the word the learner saved while reading the paragraph of Figure 2 , now carried into conversation. At the bottom are the push-to-talk button and a text box whose placeholder reads “If you can’t say it, type it; press Enter to send.” (b) The writing panel. The header reads “Writing practice: after reading a section, retell or summarize it in English; the local LLM corrects it and checks whether you used your vocabulary words.” The button below the text box submits the draft for correction.
Figure 4: LLMersion Narrator, the companion tool of Section 6.6 ; its interface is in English. (a) The input page: the learner pastes text (or uses the “Open a file” and “Open a folder” tabs for batch work), chooses an accent, voice, speed, and number of repetitions, selects the outputs (MP3 audio, MP4 read-along video, PDF of the text), and presses “Read it aloud”. (b) A frame of the generated video in its light style: the sentence currently being spoken is highlighted, and the title and paragraph counter run along the top.
Figure 5: The voice and model panel of LLMersion-1, an example of the plug-in registries of Section 6.7 (Chinese labels translated here). The header reads “Voices and models”, followed by the number of speech engines currently usable and the folder scanned for cloned voices; the buttons on the right rescan for engines and free GPU memory. The text box holds a sentence from the open document, and the button beside it plays that sentence in every checked voice for side-by-side comparison. On this machine only the Kokoro engine (tagged “local”) is installed, with American and British voices. The greyed cards below are the other registered engines: Chatterbox (tagged “clonable”, “local”, and “speed handled by the front end”), Piper (“local”), Edge neural (“online”), and the operating system’s voice (“local”); each shows, in English, the one-line installation step that would enable it.
| Commercial | Cloud LLM | LLMersion-1 | ||
|---|---|---|---|---|
| lang. app | chat | reader | (+curator) | |
| Learner’s own documents | rarely | paste-in | ✓ | ✓ |
| Synchronized read-while-listen | fixed library | — | — | ✓ |
| Inline translation | ✓ | ✓ | — | ✓ |
| In-context vocabulary capture | ✓ | — | limited | ✓ |
| Spoken conversation on own documents | scripted | cloud tiers | — | local |
Table 4: Feature comparison with the tool categories available to an individual learner. Entries are design properties drawn from public product documentation, not measured outcomes; interactive spoken feedback from a human tutor remains the quality benchmark.
| Component | Params | Memory |
|---|---|---|
| TTS (Kokoro) | 82M | 0.4 GB |
| MT (opus-mt-en-zh) | 77M | 0.3 GB |
| ASR (Whisper base/small) | 74–244M | 0.3–0.9 GB |
| Phoneme CTC | 0.3B | 1.2 GB |
| Tutor LLM (quantized) | 1–4B | 0.8–2.5 GB |
| Parsing, UI, cache | — | 0.2 GB |
Table 5: Default resident components of the reference instance ( hexgrad, 2025 ; Tiedemann and Thottingal, 2020 ; Radford and others, 2023 ; Xu et al., 2021 ; hugging-quants (Hugging Face), 2024 ) .
| Channel | 5 years | Basis |
|---|---|---|
| LLMersion-1, new floor laptop | $245 | 18 electricity |
| LLMersion-1, owned device | $18 | electricity, 60 W |
| Cloud LLM subscription | $1,200 | $20/month |
| Tutoring, 1 h/week | $2,600 | $10/h floor |
| Korea, English tutoring | 14.9M won | 248k won/month |
| China, tutoring household | RMB 42,190 | RMB 8,438/year |
Table 6: Five-year cost of one hour of daily practice, derived from cited prices ( Walmart.com, 2026 ; U.S. Energy Information Administration, 2025 ; OpenAI, 2023b ; italki, 2026 ; Statistics Korea and Ministry of Education, Republic of Korea, 2025 ; Wei, 2024 ) . Native currencies preserved; the device row is a dated retail snapshot, not a market average.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Time using the system: first session 1 week 1 month 1 year | 0 | 1 | 2 | 3 | 4 | |
| Reading fluency | ||||||
| 1 | I can read English technical documents for longer sessions without fatigue. | |||||
| 2 | My reading speed on technical English has increased. | |||||
| 3 | I rarely need to re-read a sentence to work out its structure. | |||||
| 4 | I rely on the translation less than when I started. | |||||
| Listening | ||||||
Table 7: The LLMersion-1 preliminary longitudinal self-report instrument (designed; not yet administered or validated). Mark one box per row: 0 = strongly disagree / never, 4 = strongly agree / always. Items 21–24 apply only to learners who use the night curator.