A speech-to-text workaround for Cornish
Why I tried this
I often dictate messages rather than type them because it is quicker and easier. I would like to do the same in Cornish, but there is currently no reliable Cornish speech-to-text system that I know of.
I therefore tried to build a temporary workaround using existing tools. The idea was not to make Whisper recognise Cornish directly. Instead, I wanted to see whether another tool could reconstruct the intended Cornish from Whisper’s inaccurate transcription.
Demonstration
I recorded:
Fatla genes? Pyth yw dha hanow?
Whisper treated the recording as English and produced:
“Fat again is… Pith you the Hanobe.”
I then pasted this transcription into NotebookLM, which had access to my collection of Cornish grammar resources, together with the instruction:
Please use my grammar sources to reconstruct this messy Whisper transcription as correct Standard Written Form Cornish.
NotebookLM returned:
Fatla genes? Pyth yw dha hanow?
In this example, it successfully recovered the sentence I had originally spoken.
What this shows
This is not true Cornish speech-to-text. Whisper is still interpreting Cornish sounds as English, and NotebookLM is then trying to infer the intended Cornish from that distorted output.
Even so, the experiment shows that two imperfect tools can sometimes be combined to produce a useful result. Whisper preserves enough of the sound pattern for a grammar-grounded language model to reconstruct the likely sentence.
The process also encouraged me to speak clearly and made the relationship between pronunciation, spelling and grammar very noticeable.
The limitations are substantial. It is slow, it may work less well with longer or less predictable sentences, and the output must still be checked by someone who knows Cornish. However, as a prototype, it suggests one possible interim approach while no dedicated Cornish speech-recognition model exists.
The same principle might also be worth testing for other minority languages with limited speech technology but a reasonable collection of reliable written resources.
Technical notes
The workflow uses a Mac shortcut that:
- prompts me to record some Cornish;
- runs Whisper through a Python script;
- copies the transcription;
- opens my Cornish grammar notebook in NotebookLM;
- allows me to paste in the Whisper output for reconstruction.
The process took close to a minute for this very short example.
I did not write the Python code myself. I described the workflow I wanted to ChatGPT, asked it to generate the code, then copied, tested and adapted it for use on my Mac.
