fix: remove forced language param - causes hallucination on tiny model

fix: raise no_speech_threshold to 0.95 for lenient segment filtering

fix: add segment-level logging, relax no_speech_threshold

fix: normalize audio to [-1,1] range before Whisper

fix: resample audio to 16kHz, add debug logging

fix: add python-multipart dep for file upload

Initial deploy: faster-whisper FastAPI server

initial commit