alt Hacker NewsThis was an fun read, as someone who's both a Korean BW player and a speech recognition researcher.
It's interesting to note that the original Korean transcription already has many errors, seemingly (and impressively) corrected by LLMs later on. For example, 12 안마당 빌드 (12 courtyard build) is actually 12 앞마당 빌드 (12 frontyard build), which might have been more understandable to BW players. Similarly 투에처리 빌드 (processing-at-two build? makes no sense lol) should have been transcribed 투해처리 빌드 (two-Hatchery build).
Therefore it may also be helpful to directly feed the slang dictionary into Whisper's inference process using contextual biasing. There are lots of ways to do this, but the simplest would be to increase the probability of slang words in the dictionary in the final prediction layer of Whisper by a constant factor. This is fairly easy to implement, for example by using HuggingFace's library: https://huggingface.co/docs/transformers/en/internal/generat...
Replies
I am a StarCraft fan and I have no idea what a courtyard or a frontyard is supposed to be! However I do know that the names of buildings, units, technologies, and strategies are usually heavily abbreviated in English. Perhaps the same is true in Korean? A 12 barracks build would usually just be called "12 rax", a two hatchery mutalisk build would be called "2 hatch muta", and a three hatchery hydralisk timing attack / all-in would be called "3 hatch hydra bust".
Do they actually use the Korean word for, like, tossing something to refer to the Protoss? That’s a pretty funny cross-language pun if so.
Thanks for the added context on the builds! As "foreign" BW player and fellow speech processing researcher, I agree shallow contextual biasing should help. While not difficult to implement, most generally available ASR solutions don't make it easy to use. There's a PR in ctranslate2 implementing the same feature so that it could be exposed in faster-whisper: https://github.com/OpenNMT/CTranslate2/pull/1789