Is it possible to limit autodetect to particular list of languages?

[sakhdevel]

Hi there.

First of all, thank you for this nice software. I'm we're happy with it, and I was thinking that there are some models that support many languages. But in most cases, random user will want to use basically two of them. For example, English and his own language. So is it possible to limit this after detect feature to work with these two languages so it will not be recognized as any other language?

[IvanMurzak]

This is a great feature proposition! I just was about to request the same one.

[shockme]

This is the case for me, I use usually 2-3 languages. Many models understand what I'm saying in non-English Latin languages, but translate to Russian or Chinese. I have translation disabled in settings, so this baffles me.

[ivangretsky]

Yep. A list of languages instead of just one or autodetect would be super useful (if it works). I need 2 languages like @sakhdevel - English and my native. If I choose just the native language it would still sometimes get my English includes. But if I start with an English word, it would just wait for me to say something in my native. At least this is how it works with Nemotron Streaming. If I choose autodetect it will quite often confuse my native language with similar. So trying a list of languages seems like a good idea.

[cjpais]

I believe there's been other issues and discussions on this in the past. Most models do not accept this. So it's not genuinely possible most of the time. There are a class of models that do, but largely this is not being implemented because it's just not something the models were trained for, and as a result of that, we can't do much

Again, for some models, this is potentially possible, but it's also quite possible that it's such a marginal gain on average because the models were not necessarily trained with this capability, that it may not improve performance even if we give a setting for it. And then that means that we have one more setting to maintain that actually doesn't necessarily help anyone. And that seems like a bad trade-off to me. So I agree that this is very logical and probably should be something that models are trained for, but I'm quite skeptical that things work properly end-to-end, so it's kind of just a low priority at the moment because it honestly needs deeper investigation that will take some time. And also at the end of the day this is effectively for transcribe.cpp