A Community Kept Authority While AI Helped Learners Practice Baba Malay

A small Singapore pilot offers a useful model for artificial intelligence that expands human agency without claiming human authority.

ChakapBot turned community-approved Baba Malay dictionaries, teaching materials, grammar research and human-recorded speech into a flexible practice tool. Learners could ask questions, hear pronunciations and revisit lessons on their own time, while an experienced teacher and community experts remained responsible for cultural context, source selection and validation.

The result is promising but narrow. Twenty-six people completed the six-week program, and the study measured engagement and perceived usefulness through surveys rather than objective language proficiency or long-term language vitality. The finding is therefore not that AI “saved” an endangered language. It is that a constrained system made community knowledge easier to practice while people retained authority over what it knew and how it taught.

Research references: [1]

A language tool built around community authority

Baba Malay is a critically endangered heritage language associated with Peranakan Chinese communities in Singapore and Malaysia. In Singapore, English-medium education, changing family patterns and limited intergenerational transmission have reduced opportunities for younger people to hear and use it.

ChakapBot was developed through a collaboration that included community member and linguist Nala H. Lee, Baba Malay educator Baba Kenneth Chan, heritage expert Baba Andy Gwee, community member Nyonya Josephine Tan and a Temasek Polytechnic development group led by Huiyu Zhang. The Peranakan Association Singapore supported community engagement, and Singapore’s National Heritage Board funded the work through Heritage Research Grant HRG-60.

The team did not ask a general-purpose chatbot to improvise the language from the open internet. It assembled a closed corpus from *A Baba Malay Dictionary*, the community-authored textbook *Mari Chakap Baba* and a grammatical description produced through language documentation. The authors report obtaining permission from the relevant authors or rights holders, all connected to the Peranakan community.

That provenance matters. For culturally sensitive, low-resource knowledge, the question is not only whether an AI response sounds fluent. It is who selected the sources, who consented to their use and who can correct the system.

Research references: [1] [3]

Structure came before generation

The system combined three parts: a learner-facing chatbot, a web application for recording and validating audio, and an internal AI Training Hub for managing resources and retrieval.

Much of the learning experience was deliberately structured. Dialogflow intents followed the sequence and objectives of the community textbook, supporting guided lessons, listening and translation exercises, quizzes and topic navigation. Dictionary questions retrieved entries from the digitized community dictionary.

For more open-ended questions, ChakapBot used Gemini 2.5 Flash behind a retrieval-augmented generation pipeline on Google Cloud Vertex AI. The closed corpus was divided into overlapping chunks, indexed with a multilingual embedding model and retrieved through similarity matching. The paper reports that no external data sources or APIs were connected to the generation pathway. A system prompt instructed the bot to ground answers in the approved knowledge base and acknowledge when information was unavailable. Learners also saw an AI disclaimer advising them to verify important information with trusted sources.

Those controls narrow risk; they do not eliminate it. The study did not report quantitative tests of retrieval accuracy, hallucination rates, refusal reliability or privacy performance. “Closed corpus” describes the information connected to this application’s retrieval pathway, not the training data of the underlying hosted model.

Research references: [1]

Human voices were infrastructure, not decoration

Baba Malay is learned and used through speech, so written retrieval alone would miss a central part of the language. Community voice contributors recorded pronunciations, phrases, idioms and conversational examples.

The version of record makes a strikingly large claim: it says each of two speakers produced 8,714 recordings. Those recordings were stored for human review, and a heritage expert checked pronunciation, intonation and alignment with written forms before approved audio entered the learning experience.

The exact count is worth preserving because it shows the scale reported by the authors, but the paper does not provide an independent audit of the recording database. What the described workflow establishes is clearer: people supplied the voices, determined correctness and decided which material could be reused. AI helped organize and deliver the material; it did not become the source of linguistic authority.

Research references: [1]

What the pilot found

Forty people initially signed up, and 26 completed the six-week program and both surveys. Participants ranged from 10 to 40 years old. Only three identified as ethnically Peranakan, while eight reported childhood exposure to Baba Malay. Recruitment occurred primarily through the Peranakan Association Singapore’s social channels, and completers received vouchers.

The course blended people and software. An experienced Baba Malay teacher led the first and third lessons in person; the second and fourth lessons took place mainly in the app. Learners could also use the app for review during the in-person weeks.

Most completers reported using ChakapBot for at least 15 minutes a week; seven reported less than 15 minutes. About half rated their perceived improvement at four or five on the study’s scale, and none selected zero. Participants highlighted flexibility, audio support, spontaneous questions and a low-pressure space for independent practice.

Those are self-reports, not measured learning gains. Participants attributed more of their progress to the in-person classes than to the chatbot, and the authors describe the two modes as complementary. Learners also identified gaps: no speech-recognition or “speak-to-check” function, uncertainty about accuracy and limited conversational depth.

Research references: [1]

The agency lesson is in the division of labor

The most useful reading of ChakapBot is not “AI replaces a language teacher.” It is that a community used AI as delivery infrastructure for knowledge it had selected, licensed, recorded and reviewed.

That arrangement supported several forms of agency:

• **Learner agency:** people could practice when they had time, revisit material and ask questions without the pressure of speaking before a class.
• **Teacher agency:** structured digital practice extended an existing syllabus instead of discarding it.
• **Community agency:** community members controlled source materials, voices, validation and boundaries on what the system should say.
• **Institutional agency:** the internal training hub made corpus changes and governance visible to the project team instead of burying them inside an opaque consumer chatbot.

What remains unproven

This was a six-week pilot with 26 completers, no control group and a volunteer sample recruited largely through a community association’s social channels. Fourteen of 40 sign-ups did not complete the program and surveys. Only three completers identified as Peranakan, limiting conclusions about the heritage community most directly affected.

The study did not include a standardized proficiency test, long-term retention measure or direct measure of language vitality. Survey responses can describe experience and perceived usefulness, but they cannot establish that the chatbot caused learning or reversed language decline.

The Temasek Polytechnic Ethics Committee approved the study on October 21, 2024, under code SR35-IIT-2024_ER. The paper says informed consent was obtained from all participants and that no personally identifiable information was included in the analysis or reporting. Because the sample included a participant in the 7–12 age band, the absence of a separate published description of guardian consent and minor assent is a reporting limitation; it is not evidence that those procedures did not occur.

Participant data are available only on request because of privacy and ethical restrictions. The authors declare no conflict of interest, and the project was funded by Singapore’s National Heritage Board through grant HRG-60.

A larger evaluation would need objective pre/post assessment, longer follow-up, clearer dropout analysis and stronger representation from the heritage community. Technical evaluation should also test retrieval accuracy, refusal behavior, speech feedback, privacy safeguards and performance as the corpus grows.

Research references: [1] [3]

A restrained conclusion

ChakapBot is an early pilot, not a verdict on AI and language preservation. Its strongest contribution is a design principle: for culturally sensitive, low-resource knowledge, the community should be the authority and the model should be a bounded interface.

That principle extends beyond language learning. In public services, accessibility, community archives and science communication, AI can help people navigate knowledge without claiming to own it. The measure of success is not how human the system appears. It is whether people gain meaningful room to learn, decide, create and participate while retaining control over the sources and values that shape the system.