Measured with the ort-web-under-Bun harness: no reliable speed difference on
clean, language-matched audio. On out-of-distribution audio (wrong/undetected
language), the multilingual model's extra forced task token measurably
changes how long its repetition-loop hallucination runs before EOS (434 vs 67
tokens on the da_festuge fixture), which is where the reported slowness comes
from. Both models resolve to identical, merged, KV-cache-enabled decoders at
the same dtype — no catalog/config fix applies here. Documents the mechanism
and flags a future no_repeat_ngram_size/repetition_penalty mitigation as the
actionable follow-up if this recurs.
Lucas Winther
7ab41f4299