Translating game-specific vocab

The problem with gamer language

Here's the thing with livestreaming games - a lot of the words you use don't appear in normal dictionaries, or if they do they mean something different than you are using them. Like any community, gamers and streamers develop their own jargon and use of language, and this can be difficult to parse for standard vanilla systems.

BabelStreamer is optimised for gamers and streamers, so how do we tackle this?

Start at the beginning - speech recognition

It all starts with recognising what the streamer is trying to say in the first place. If you're using game specific vocab, or even personal catch phrases, the recognition software might just not get it. So BabelStreamer has a two-tiered approach, based on biasing the recogniser to specific vocab. Full disclosure: this is a bias and not a guarantee, and no recogniser is going to get it 100% right.

This is why it is important to tune the recogniser to your own speech patterns - the better the source the better the translation. Also you might want to try a larger recognition model than the standard, if your system has enough headroom.

Specific game vocab

We keep lists of vocab that are specific to specific games. So for example Minecraft we have "Minecraft", "Creeper", "Enderman" etc. This is downloaded by the plugin on startup so that we can regularly update the lists and games without you having to load a new plugin version. You just select a game in the plugin and the rest is on autopilot.

Personal vocab

Last but not least the plugin allows you to enter a personal vocab file which will bias the recogniser to your own specific vocab.

To translate or not to translate

Once the text has been recognised it is sent off to the server where it will be translated into the user-requested languages and distributed. At this stage we have to do more work on the game vocab.

When not to translate

Game vocab is generally not translated but just passed on unchanged (at least with Latin scripts). The standard vocab - general gaming terms such as "oneshot" or "nerfed" - is always protected in this fashion, as is any selected game vocab.

This produces some challenges of its own as the translators can't be told to do this, so we have to preprocess the text and handle the grammar ourselves (yes - I'm looking at you German!). This is why we limit the source languages for translation - each language pair needs to be tested individually.

When to translate

Of course anything else is just translated normally, with one extra twist. Translators have a habit of mistranslating some common gaming phrases. The translator keeps a list of these common mistranslations, which we update over time, and fixes them.

End of the story and what comes next

All of the above is designed to give viewers the most natural feeling translation that we can currently offer. The one tweak in the pipeline is that we are working on a service to tune the recognition model to a specific speaker's voice, to make the recognition even more accurate.