I understand it's implemented in C, but did you look into why the cache detection didn't work?

I am not entirely sure at the moment, but the changes in this PR are based on two key observations:

• The vosk Model may attempt to download a model when the model directory is empty due to an incorrect model_path being used. We want to make sure that we use VOSK_MODEL_PATH as the path.
• When checking the LocalStack cache, we only verify whether the model path exists. It's possible that the directory exists but doesn't actually contain any model files.

I have added a log to check if we use a cached model and log its path.

We can take as a next step to verify the checksum if the tests remain flaky, let me know thoughts on it?

The vosk Model may attempt to download a model when the model directory is empty due to an incorrect model_path being used. We want to make sure that we use VOSK_MODEL_PATH as the path.

If model path is now passed as part of Model instantiation, is it still necessary to set VOSK_MODEL_PATH?

When checking the LocalStack cache, we only verify whether the model path exists. It's possible that the directory exists but doesn't actually contain any model files.

Any idea what would lead to this situation, where empty directories are created?

I think our CI networks are pretty resilient and shouldn't cause datastream to be corrupted, otherwise this would pop up in several other tests that we have. Adding checksum verification is probably unnecessary, I suspect the root cause is something else.

Yes absolutely, thanks for pointing that out 🙌 I have removed VOSK_MODEL_PATH in b6c7b46.

One of the possible reasons of failures could be that we are setting up cache directory from the environment variable. Which needs to be set before vosk module is imported. We currently import vosk in different locations - which could be the possible issue of cache path not being set.