Strange behavior between CPU to GPU: incorrect language identified + translation

[hmf]

I experiment with the project using the git repo. My setup is as follows:

Linux Ubuntu
GPU: Quadro P1000 with 4GiB
CUDA Version: 12.4
    NVIDIA-SMI 550.120
    Driver Version: 550.120
Using a fork from this repository
Using a branch from the original master
Created and using a virtual environment
Using Python 3.12.3
Installed all libraries from all 3 requirements files
Executed the test as indicated in the CONTRIBUTING.md file

As explained here my Quadro 1000 has compute capability 6.1. I tried this example of the README, which uses the large-v3 model. I changed it slightly to use the "./tests/data/physicsworks.wav" sample and use the CPU (int8). This is what I get:

Detected language 'en' with probability 0.547627
[0.00s -> 6.62s]  Now I want to return to the conservation of mechanical energy.
[6.62s -> 12.32s]  I have here a pendulum. I have an object that weighs 15 kilograms,
[12.32s -> 15.66s]  and I can lift it up one meter, which I have done now.
[15.66s -> 19.94s]  That means I've done work. Mgh is the work I have done, believe me.
[19.94s -> 23.00s]  I've increased the potential energy of this object.
[23.00s -> 29.88s]  Fifteen times ten is about 150 joules. If I let it fall, then that will be converted

That seems to be working pretty well, but it is a little slow "heating" up my CPU. So I changed this to (float32 causes out of memory):

model = WhisperModel(model_size, device="cuda", compute_type="int8")

and get:

Detected language 'nl' with probability 0.561300
[0.00s -> 4.82s]  Nu wil ik terug naar de conservatie van mechanische energie.
[6.30s -> 8.98s]  Ik heb hier een pendule.
[9.82s -> 14.82s]  Ik heb een object dat 15 kilogram weegt en ik kan het op 1 meter lopen, wat ik nu heb gedaan.
[15.48s -> 16.72s]  Dat betekent dat ik werk heb gedaan.
[17.54s -> 19.68s]  Mgh is het werk dat ik heb gedaan, geloof me.
[19.90s -> 22.20s]  Ik heb de potentieelende energie van dit object vergroten.

This is baffling because the spoken language is english. Granted the speaker does have a dutch accent, but the model seems to be translating and not transcribing.

Can anyone explain this behavior?

TIA

[hmf]

Using "turbo" as is shown in the second README example consistently transcribes english.