Learning a language has become easier with the help of language learning apps like Duolingo and translation tools like Google Translate. However, these apps are not effective for speakers of Indigenous languages in regions like North America or New Zealand, where efforts were made to replace local languages with European languages. Michael Running Wolf, a PhD student studying natural language processing at McGill University, realized this when he was working on speech recognition algorithms for Amazon’s Alexa. He grew up on the Northern Cheyenne Indian Reservation and spoke the Cheyenne language with his grandmother. He found that Alexa’s algorithms would not work for his tribe’s language because many North American Indigenous languages have different structures from English or Mandarin, which are the main languages these algorithms are designed for.
One challenge is that many Indigenous languages are polysynthetic, meaning that words change form depending on the context. For example, in English, we would say “the full green cup is mine,” but in a polysynthetic language, this phrase could be expressed with a single word. Teaching speech recognition software like Alexa to recognize these words would require a new algorithm. Additionally, there are limited data sets available for most Indigenous languages, making it difficult to train AI models. Unlike English, which has tens of thousands of hours of recorded speech available for training, Indigenous languages often have only a few hours of recorded speech.
However, there have been promising developments in building language models for Indigenous languages. Keoni Mahelona, the chief technical officer of Te Hiku Media, worked with the Māori community in New Zealand to bootstrap a speech-recognition model using only 310 hours of te reo, the Māori language. This demonstrates that it is possible to build a language model with relatively limited data. Mahelona’s team is now working on Papa Reo, a platform for te reo speech-recognition apps, making it easier for New Zealanders to interact with their devices in te reo.
Running Wolf has a different goal. He wants to develop AI systems that can help people practice speaking their tribes’ languages. He envisions children using AI to supplement their language learning, with the AI recognizing mispronunciations and providing gentle corrections. Working with his wife, Caroline Running Wolf, who is a PhD student at the University of British Columbia, they are also designing augmented reality (AR) experiences for learning native languages in context. For example, they are creating an AR game for the Kwagu’ł community on North Vancouver Island, where players gather materials for a traditional potlatch feast in a virtual recreation of their ancestral lands. This immersive experience not only teaches the language but also cultural protocols.
While there is potential for AI to assist in preserving Indigenous languages, there are also important considerations. Mahelona urges large corporations not to monetize Indigenous languages and instead work with Indigenous-run organizations for access through licensing or other agreements. The historical context of colonization adds complexity to these discussions, as Indigenous languages were suppressed by governments. Running Wolf also emphasizes the importance of considering the economic development of Indigenous communities when utilizing their data.
In conclusion, AI holds promise for preserving Indigenous languages, but it requires addressing the unique structures of these languages and the limited data available. Collaboration with Indigenous communities and respecting their interests and economic development are crucial in this endeavor.
