Indigenous ASR - recognising more than speech

Speech recognition technologies are rapidly evolving and improving, but the technologies only work for mainstream languages. One primary challenge for indigenous communities wanting to use these tools is obtaining enough digitised, labeled data. We demonstrate how Te Hiku Media overcame this barrier by collecting more than 300 hours of labeled speech corpus in just 10 days. This enabled us to build the first automatic speech recogniser (ASR) for te reo Māori using DeepSpeech. We now use ASR to accelerate the transcription of native speaker archives. Our journey is one of community, trust, and sovereignty.