Reclaiming Voices with Small AI: Revitalizing Lost Mother Tongues through Community-Based Language Models

Countering the Flattening Effect of AI 

We often hear about the powerful impact of large AI language models. However, for many marginalized languages, these models can pose more of a threat than a benefit. Large models tend to flatten voices, identities, cultures, and linguistic nuances (Gafour & Chehri, 2025). Why does this happen? Because marginalized languages are often underrepresented in digital data, and in some cases have even ceased to function in professional or even everyday contexts. As a result, AI systems have little reliable data from which to learn.

But what if we flipped this challenge into an opportunity? What if small language models, designed for targeted, community-specific language input, could help preserve and revitalize languages at risk of disappearing? Could they also serve as a tool to foster teachers’ critical language awareness?

A Personal Journey 

My 91-year-old father has long spoken of writing his memoir for many years, yet he has remained stuck on chapter one. He types in Chinese on a Word document, but he has a habit of repeating the cycle of writing and erasing. While I cannot second-guess what is happening in his mind, it sometimes seems that he is enduring a cultural-neurological tension that compels him to relive through the cultural and linguistic erasure he experienced as a Taiwanese living under a Chinese-dominated regime in his early years.

Last year, when I visited my parents, I decided that I wanted to help my father preserve his legacy by supporting him with his memoir.

As AI became widely accessible, I thought I could simply guide him to speak into a system that would transcribe his stories in Taiwanese, Hokkien—a local “dialect” spoken by a large portion of people in Taiwan. Although there has been a recent revival, Taiwanese has historically been marginalized in public spaces under a Chinese-language dominant ideology. Moreover, as a language, Taiwanese has no standardized written system. In the past, an oral history project like this required recording audio and manually transcribing it phonetically into Chinese—a time-consuming and labor-intensive process. With the advent of AI, however, I imagined a game-changer had arrived.

The reality, however, proved more complicated than I had envisioned. I was not sure if AI could fully recognize Taiwan’s complex linguistic reality. My parents speak Taiwanese (Hokkien) in everyday life. Although they can speak Mandarin and read and write Chinese characters, their primary language of communication is Taiwanese. In daily conversation, they frequently switch between Mandarin and Taiwanese. To support my father’s project, I faced a challenge: I am fluent in Mandarin and literate in Chinese, but I understand Taiwanese only at a rudimentary level. I cannot speak it fluently, let alone write or type it, especially given its lack of a standardized written form. As a result, I had to rely on AI to recognize and navigate this back-and-forth linguistic complexity.

Over time, AI may enable us to transcribe all spoken languages directly. Yet an important question remains: can AI capture not only the words themselves, but also the way they are spoken—the cadence, pronunciation, and emotion? These elements are essential components of the historical record, and without them, the richness of oral histories cannot be fully preserved.

Launching a Small Language Model (SLM) Project 

With these questions in mind, I began experimenting with the idea of using a small language model to document my father’s storytelling and memoir writing.

Although the literature on small language models is growing, it tends to be highly technical. My interest lies primarily in their practical application, specifically, how they can be used to document and annotate language when only small datasets exist due to marginalization.

In this context, I understand a small language model (SLM) as a custom AI system built for a specific purpose (Fraisse et al., 2019; Koc, 2025). Unlike large language models trained on massive datasets for general tasks, an SLM focuses on a narrow linguistic domain, such as a specific language, dialect, or community. Because it is trained on a smaller, carefully curated set of examples—such as recordings of elders speaking a local language—it can better capture the unique vocabulary, pronunciation patterns, linguistic structures, and stories that might otherwise be lost in large-scale AI systems.

With the assistance of AI tools, I conceptualized the following workflow for a small language model project:

1. Record the living language Capture oral histories or life stories in audio format.

2. Phonetic transcription Convert the speech into a phonetic representation first, as romanization is still widely recognized internationally as a way to represent sound. In some languages, such as Chinese, pronunciation can vary significantly across local dialects. This step preserves those pronunciation nuances—even if they are nonstandard—and captures the “living language” of the speaker.

3. Transcribe into a target language For my project, the next step is transcription into Chinese so my father and others who are unfamiliar with the romanization system can review and verify accuracy.

4. Translate for broader access Optional translation into English allows the stories to reach a wider audience.

5. Reflect and analyze Compare the original speech with the transcribed and translated versions. Which aspects of meaning, cultural knowledge, and linguistic nuance were preserved? Which were lost?

This final step is particularly important. The goal of this language project is not merely to preserve information, but also to preserve the language as a lived experience. It also documents our insights as translingual individuals navigating across languages and cultures.

A Pilot Example 

As an initial experiment, I decided to use a recorded conversation between my parents as a pilot study. In this conversation, my mother explains why she does not yet need a hearing aid.

The following demonstrates a step-by-step record–transcribe–annotate process. You can listen to the original audio here: https://drive.google.com/file/d/1dfsn-GSx7gTX_pGkNV5z7spUgLH2-XHc/view?usp=sharing

1.     Uploaded the audio file to Google Gemini (Free version) to produce a phonetic transcription 

Romanized Audio Transcription of Speech

2. Generated a Chinese transcription and an English translation

Chinese Transcription 

A: 或者是電腦仔，抑是手機仔用一睏仔。 

B: 靠背音啦。 

A: 倒去音，倒嗓聲。 

A: 啊，若是……彼我若咧看電腦仔。 

A: 電腦我若無鬥彼個（助聽器）喔。 

A: 聲音按呢，暗暗，毋是講一成清楚。 

A: 我就鬥彼粒耳機（助聽器）呀。 

A: 倒彼個聲音就直接咧講話啦，就聽有。 

A: 啊若咧電腦按呢，就按呢，抑是咧講話就聽真清楚，就一直看字一直看字。 

A: 袂曉真清楚。 

A: 總是真大聲嘛是，喔，出力。 

A: 啊這馬我就愛靠耳機呀。 

A: 我彼個耳機害去，賰一爿也咧聽。 

A: 欲加聽一下啦。 

A: 欲買一個預備的啊。 

A: 遐無講真有欲（聽清楚）啦。 

A: 虛空仔，當時仔，咱去運動，當時仔老師咧講啥物，按呢有聽到有無聽到按呢.

A: 無法度逐句攏聽甲真清楚。 

B: 喔。 

A: 眾人衰退啊。 

A: 佮這馬去行路，我去行路按呢看車，攏嘛是，若像罩茫仔。 

B: 正常啦，八十九歲矣，你毋是六十九歲。

B: 六十九歲就已經老矣。 

A: 著啊。 

B: 凡若是五十以上就老翁老婆矣。 

A: 自底的跤骨 ，啊剪仔姨啦 ，完仔姨啦，啥物人姨，伊看著講「喔，較緊喔」，當初家己這馬無啥欲看人。

A: 遮樓跤樓跤路个問，問講「啊恁查某囡仔轉來矣？」

A: 我嘛聽，知伊咧講話，聽無真清楚。 

3. English Translation 

A: Or if I’m using the computer or my phone for a little while.

B: It’s because of the background noise.

C: The sound echoes back, and my voice starts to give out.

A: And, if... like when I’m watching things on the computer.

A: If I don't put on that thing (the hearing aid) for the computer...

A: The sound is like... muffled; it’s not even ten percent clear.

A: So I have to put on that hearing aid.

A: Once that's in, the sound comes through directly when people are talking, and I can hear.

A: When I'm at the computer like that, or when people are talking, I can hear very clearly, and I just keep reading the captions/text over and over.

A: Otherwise, I can’t understand it clearly at all.

A: Even when it’s very loud, it’s still... ugh, such an effort.

A: So now, I have to rely on the hearing aid.

A: Mine broke, so I only have one side left to listen with.

A: I have to try harder to hear.

A: I’m thinking of buying a spare one.

A: It’s not that I’m being stubborn about it...

A: It’s just that sometimes when we go exercise, and the teacher is saying something, I catch some parts but miss others.

A: I just can't hear every single sentence clearly anymore.

B: Oh, I see.

A: Everyone’s health declines eventually.

A: Like now when I go walking—I go out walking and look at the cars, and it’s all just... like a blur.

B: That’s normal. You’re eighty-nine years old, not sixty-nine!

B: Even at sixty-nine, you're already considered old.

A: That’s true.

B: Generally, once you’re over fifty, you’re already an old man or an old woman.

A: My legs used to be so strong... Auntie Jian, Auntie Wan, and the others... when they see me, they say, "Oh, walk faster!" But honestly, these days, I don't really want to see anyone.

A: People downstairs on the street ask me, "Oh, is your daughter back?"

A: I can hear them, and I know they are talking, but I can't quite make out what they're saying.

4. Analyzing Inaccuracies and Noting Lost Cultural and Linguistic Elements 

While the translation captures the general meaning and overall content of my mother’s speech, there are some significant errors. Below, I have highlighted several discrepancies between the transcribed versions and the original speech:

1)    Through the phonetic transcription process, I became increasingly aware of both the possibilities and limitations of AI. For example, the word for “computer” (電腦) is romanized in Taiwanese as tiān-náu, which does not exactly match how my mother pronounces it. There are two possibilities: either the speaker pronounced the word differently and the AI automatically corrected it to a standardized pronunciation, or the system lacked data, such as tone markings, and could not represent the speaker’s words accurately. Either way, this points to the inherent limitations of AI in capturing tonal nuances digitally.

2)    In another instance, my mother explained that a computer or iPhone can amplify sound. When AI transcribed the phrase into Chinese characters, it produced a word that was phonetically similar but semantically incorrect. The correct term should have been 擴音 (amplify sound), yet the system generated 背音, which conveys a completely different meaning.

3)    There were also moments of translanguaging. For words like “computer” or “hearing aid,” my mother naturally switched into Mandarin, since these are imported technological terms. However, this linguistic shift was not fully captured in either the Chinese or English transcription.

4)    Another example involved the name “Eti,” short for Etiyawati, the Indonesian domestic helper who works with my parents. AI failed to recognize or accurately transcribe her name. In the transcript, the reference to her disappeared entirely, effectively erasing her presence and identity from the story.

Because of my multilingual background, I could recognize how my mother’s language was being truncated through this process. The richness, musicality, and nuance of her speech were not fully preserved, highlighting the limitations of AI in capturing not just words, but the cultural and linguistic texture embedded within them.

Classroom Implications 

Small language models (SLMs) are not only useful for personal projects; they also hold significant educational potential. In particular, they can function as part of a decolonizing approach to language education by challenging the dominance of major languages and creating space for marginalized linguistic practices.

For students, especially multilingual or translingual learners, this process can provide opportunities to:

• Reconnect with heritage languages that may have been marginalized or excluded in school settings

  
  

• Analyze phonetic and semantic nuances across languages, noticing how meaning shifts through transcription and translation

  
  

• Engage critically with language preservation, including how digital tools represent—or fail to represent—their languages

For teachers, SLMs can serve as a powerful pedagogical tool. They allow educators to facilitate student-driven projects that document community knowledge, oral histories, and linguistic diversity. At the same time, such projects highlight the translanguaging insights that multilingual learners bring into the classroom.

In addition, engaging with SLM-based projects can help teachers develop critical language awareness. As teachers work through the process of recording, transcribing, and analyzing language data, they begin to see the risks of digital misrepresentation—moments when AI systems either distort students’ languages or fail to recognize them altogether. These experiences can deepen teachers’ understanding of how language, technology, and power intersect in educational contexts.

When I presented this idea in a teacher education course, it resonated with several prospective teachers. One student shared a related project:

“I’m personally working on a docuseries interviewing an older colleague from my job who is also a fellow musician. He’s about 75 and often shares one- to three-hour stories when he comes in. I was thinking of using speech-to-text software to transcribe and edit subtitles.”

Stories like this align closely with a community asset mapping framework, where individuals use digital tools—including AI—to bring community voices and histories into the classroom.

When used thoughtfully, AI can help bridge generations, preserve marginalized languages, and make personal histories accessible to broader audiences. More importantly, it can encourage learners and educators to see language not merely as a tool for acquiring dominant languages or meeting neoliberal aspirations, but as a living repository of identity, culture, and collective memory.

References

Fraisse, A., Zhang, Z., Zhai, A., Jenn, R., Fisher Fishkin, S., Zweigenbaum, P., Favier, L., & Mustafa El Hadi, W. (2019). A sustainable and open access knowledge organization model to preserve cultural heritage and language diversity. Information, 10(10), 303.

Gafour, Z. I., & Chehri, Y. (2025). Digital Doppelgängers: Navigating Identity, Ethics, and Voice in the Age of AI Personas. Journal of Languages and Translation, 5(3), 01-14.

Koc, V. (2025). Generative AI and large language models in language preservation: Opportunities and challenges. arXiv preprint arXiv:2501.11496.

Dr. Ching-Ching Lin is Founder and Managing Director of the Global Diversity and Inclusion Lab, a faculty member at Adelphi University, and a U.S. Department of State English Language Specialist. She is dedicated to advancing culturally responsive teaching, with a focus on diversity, inclusion, and equity in education. This post was originally published April 8, 2026 and is cross-posted here with her permission. You can reach Dr. Lin by email here.

Explore Dr. Lin's Work in Community-Centered, Equity-Driven Education Book: Centering Multilingual Learners in School Curriculum through Community Asset Mapping This volume examines the theoretical foundations of community asset mapping and offers rich case studies and classroom examples across diverse educational contexts.

Online Course (Udemy) Empowering Education Through Community Asset Mapping A practice-oriented course designed for educators seeking to center students’ lived experiences and community knowledge in curriculum design.

Professional Services Learn more about the educational consulting, professional development, and research initiatives she leads through the Global Diversity & Inclusion Lab.

-------------------

Got something that needs to be heard? We'll get it said and read on the Worthy Educator blog! Email it to walter@theworthyeducator.com