Lower default KENLM_ALPHA 0.5→0.2, BETA 1.5→1.0

#5
by chirag18 - opened

First end-to-end test showed alpha=0.5 was too aggressive — LM was
overriding acoustic evidence at word boundaries and dropping leading
characters (Severe→evere, soft tissue→ft issue, etc.). Conservative
alpha=0.2 keeps word boundaries intact while still benefiting from
LM on word-choice errors. Tunable via env var without rebuild.

deepakkaura changed pull request status to merged

Sign up or log in to comment