Spaces:

---

deepakkaura

/

medasr-server

like 0

Sleeping

Three changes that together enable shallow-fusion decoding with a domain LM trained on 731K in-domain radiology reports (~111M words):

• server.py: _build_decoder() loads /app/radiology.bin if present and passes it to build_ctcdecoder with alpha=0.5, beta=1.5. Falls back to non-LM beam search if missing. /health reports kenlm_loaded.
• Dockerfile: fetches radiology.bin (240 MB) from chirag18/radiology-stt-assets at build time.
• requirements.txt: adds pypi-kenlm.

Offline sanity check on the LM passed — correctly favors 'borderline cardiomegaly' over the common ASR misrecognition 'remaining cardiomegaly' (-1.64 vs -2.21 log-prob/word). The LM should help on word-choice errors that hotwords alone can't fix.