Whisper transcription in Spanish

OS (e.g. Win10): MacOS Sonoma 14.0
PsychoPy version (e.g. 1.84.x): 2024.1.5
**Standard Standalone? Yes
What are you trying to achieve?:
I am trying to have participants record sentences in Spanish (one sentence per trial) and have PsychoPy transcribe the content of each trial. Speakers are non-native speakers of different levels of proficiency as well as native speakers. All sentences contain one non-word that follows Spanish phonotactics. I’ve created a minimal example (attached).

What did you try to make it work?:
In builder, I’ve set up the mic like so. It doesn’t ask to specify language. I’m used to–from the Whisper local app on my computer–specifying both model size AND language.

I also kicked it over to Coder to try from there and see what it looked like. I was surprised to see that it specifies language, but I didn’t see a model specified anywhere in there.

From Coder, I tried running with “en-US”, “es-US”, “es-MX”, and “es”, and I specified expected words (both real and non-word items).

What specifically went wrong when you tried that?:
The experiment runs successfully in all cases above, but the transcriptions are dreadful.

When the sentences recorded were “Paula tiene una caturra roja. Antonio tiene un falusdán marrón.”, the best I could get was specifying ‘es-US’: '“Pala tiene una katura roja, and tonio tiene un falustan maron.”

I was prepared for the non-word items (caturra, falusdán) to give it some trouble, but the fact that it can’t get the two common names is surprising.

Running Whisper on the same file from the app gives me this, which is great:

I haven’t seen much on this here in the forums. Any guidance on getting better results in a language other than English either from Builder or Coder? I’ll attach the experiment as .psyexp and as .py here. Thanks for any help!

test.psyexp (10.9 KB)
test.py (24.6 KB)