Sitting across from a laptop and microphone in a community center in the Canadian Arctic, an elder speaks slowly in Inuktitut. The researcher next to her is not a linguist in the conventional sense. With the help of the elder’s words, her perseverance, and several months of recordings, the computer scientist is creating a speech recognition model that has been trained on a language that the majority of people have never heard. The elder has knowledge that cannot be found in any database. Whether the technology can record enough of it in time is the question.
There are now about 7,000 languages spoken on Earth. According to linguists, over 40% of them are endangered, which means that they have so few speakers and so little transmission to future generations that they could practically go extinct in a few decades. A communication mechanism is not the only thing lost when a language dies. It brings with it particular classifications of geography, time, family dynamics, ecological knowledge, and cultural memory that are difficult to translate into other categories. The loss is truly irrevocable.
AI won’t be able to resolve this issue. However, it’s altering the pace at which linguists may work, and in certain instances, it’s making preservation initiatives possible that would otherwise be too labor-intensive or too slow to finish before the last fluent speakers disappear. The same underlying technology that enables voice assistants, Automatic Speech Recognition technologies, can be modified to transcribe recorded oral histories, turning spoken words into text that can be searched, examined, and stored. This is a method of producing a textual record that did not previously exist for languages without a formal written form, and it does it more quickly than human transcription alone could.
Machine learning’s ability to recognize patterns is very helpful for old and broken scripts. AI has been used to uncover structural regularities in Mayan glyphs and other imperfectly known writing systems, as well as to help analyze Linear B, the earliest deciphered form of Greek. These findings are then further investigated by human experts. The technology can recognize combinations and frequencies that take a human scholar weeks or months to notice in a matter of hours, but it cannot decipher these scripts on its own—the cultural and historical information needed to understand meaning still comes from human expertise.
Researchers can upload and access recordings, text samples, and linguistic data for languages without a central repository from anywhere in the world thanks to Google’s Endangered Languages Project. A alternative strategy is used by the Woolaroo app, which uses image recognition to teach minority languages through commonplace things. This useful consumer-facing tool lowers the barrier to casual involvement with languages that most people have never experienced. Both increase access in ways that are important for creating the community interest that supports revitalization efforts, but neither is a preservation solution in and of itself.
The real problem is that endangered languages are by definition low-resource, and standard machine learning requires large datasets to operate consistently. A model that was trained on millions of hours of English audio has good generalization. A model that was trained using forty hours of Inuktitut recordings lacks the same starting point. The margin for error in cultural preservation work is not something that experts take lightly, and although researchers have developed methods for low-resource language processing, such as transfer learning that borrows structural knowledge from related languages, the results are less reliable than they would be for well-documented languages.
Speaking with the linguists working in this field gives me the impression that the technology is both genuinely helpful and honestly inadequate on its own. It can effectively speed up and scale tasks that human researchers are already proficient in, such as transcription, pattern recognition, dictionary construction, and the development of instructional tools. It cannot, however, take the place of a researcher’s interaction with a community or a native speaker’s assessment of whether a generated sentence accurately reflects the way language functions. The most effective initiatives view AI as a tool that the community may use instead of a solution that is provided to them.
