Keep structured date extraction opt-in

Reject inferred day-level structured dates

Guard structured dates against inferred parts

Add structured date extractor

Add experimental OpenAI date sidecar endpoint

Add calendar-month hotwords to test spoken-date recognition

Preserve \n through whitespace cleanup so "new line" survives strip

Fix NameError on CORPUS_HOTWORDS

Add 86 corpus-mined hotwords

Always re-download LM on startup

Add N-best LM rescoring path

Truly bypass LM when KENLM_ALPHA<=0

Lower default KENLM_ALPHA 0.5→0.2, BETA 1.5→1.0 (#5)

Use Python 3.10 (pypi-kenlm incompatible with 3.11) (#4)

Add C++ build deps so pypi-kenlm can compile in slim image (#3)

Download KenLM at server startup, not Docker build (#2)

Enable KenLM domain-LM shallow fusion at decode time (#1)

Run model fp32 on GPU (fp16 produced degenerate logits)

Upcast fp16 logits to fp32 before pyctcdecode (fixes empty-beam error)

Use GPU when available (fp16 + autocast for T4)

Tune for speed: beam_width 8->4, hotword_weight 10->5

Fix IndentationError: collapse duplicated regex lines

Debug log + handle SentencePiece-prefixed brace tokens

Convert MedASR </s> sentence-boundary token to period

Mark pad token as CTC blank for pyctcdecode (fix off-by-one)

Fix decoder vocab-size mismatch (use model.config.vocab_size)

Add pyctcdecode for beam-search decoding

Beam search + radiology hotwords + /transcribe-pcm endpoint

Upload server.py with huggingface_hub

Upload server.py with huggingface_hub

Upload server.py with huggingface_hub

Upload folder using huggingface_hub

initial commit