Te Hiku Media decided to fine-tune Whisper using data they manage under their Kaitiakitanga License, treating data as a respected entity. The fine-tuned model performed significantly better, even providing punctuation and more accurate verbatim transcriptions. Despite these improvements, the organization maintains that they would not use such a model in production due to ethical implications and potential harm to indigenous languages. They also highlighted the need for transparency in data sourcing and the importance of respecting data in AI model development.
Key takeaways:
- OpenAI's Whisper, a speech recognition model, has been criticized for its potential to infringe on copyright and appropriate cultures, particularly in the context of indigenous languages.
- Te Hiku Media, a Māori organization in New Zealand, has expressed concerns about the use of indigenous languages in Whisper, questioning the source of the data and the potential harm of non-indigenous groups building tools like Whisper.
- Despite these concerns, Te Hiku Media conducted an experiment to fine-tune Whisper for te reo Māori and NZ English, achieving a significant improvement in word error rates.
- However, the organization maintains that it would not use a model like Whisper in production due to ethical concerns, emphasizing the importance of respecting data and the potential harm of extractive data practices.