Implemented LocalWhisperClient wrapper for local Whisper as ground-truth STT validator. ## Changes Made - **New file: src/librechat_audio/models/whisper.py** - LocalWhisperClient class supporting all Whisper model sizes (tiny through turbo) - **New file: tests/test_local_whisper.py** - 25 comprehensive tests for accuracy validation - **Updated pyproject.toml** - Added 'whisper' optional dependency group - **Updated exports** - LocalWhisperClient accessible from main package ## Verification All 25 new tests pass. Turbo model achieves 0% WER on test sentences, exceeding the >95% accuracy QA requirement. ## Usage ```python from librechat_audio import LocalWhisperClient, whisper_available if whisper_available(): client = LocalWhisperClient(model='turbo') text = client.transcribe('audio.wav', language='en') ```