Why this repo got low character error rates compared to the SYSTRAN model?

[snwefly]

I accidently found your repo and precision of your application is higher than systran faster-whisper large-v2 model, but you're also using systran models. what's exactly the reason behind this?

Thanks for the wonderful software.

[Purfview]

Here I've implemented various experimental options enabled by default to do generally a bit better transcriptions.