Hi everyone,
I’m part of a low-resource language community, and I’ve been truly impressed by how well GPT models handle Faroese, even though it wasn’t explicitly trained on it. This gives me hope that OpenAI might be open to expanding its support for Faroese in other areas as well.
Faroese has some high-quality ASR models available, with one on Hugging Face achieving a WER of 7% (self-reported on a test dataset). I’ve launched an app called VoisIT that relies on this model to transcribe Faroese speech to text. However, hosting this model on Hugging Face’s inference endpoints 24/7 is quite costly for an individual developer.
Given this, I wonder if OpenAI would be interested in adding Faroese to their Whisper model, allowing transcription to be billed on a per-token basis through OpenAI’s API.
My questions are:
- How can I approach OpenAI to explore the possibility of integrating Faroese into their Whisper model?
- If this isn’t feasible, could someone guide me on how to combine the existing Whisper model with the Faroese ASR model? I’m new to ASR training but have experience with TTS, and I’m keen to expand my app to support other languages using a custom-trained model.
For reference, the Faroese model I’m using is available here: [Whisper Large Faroese ASR Model]
Name: carlosdanielhernandezmena/whisper-large-faroese-8k-steps-100h.
Thanks for any insights or guidance!
PS. If OpenAI rather do training based on the Faroese datasets, there are two great open source datasets: Search for “Ravnur BLARK” and “Ravnursson” which are both linked to from the University of the Faroe Islands’ website. There is around 100 hours of audio with transcripts in that one.
Best regards,
Andras Eliassen