feat(ui): stabilize metrics state and expand language samples
Ping Long
d4f1e7c
- Implemented hardware-aware Performance Metrics (EST. VRAM for WebGPU, TTFT for vLLM, LATENCY for HF-GPU).
- Fixed stale engine state retention in Transcript and Metrics components.
- Removed redundant 'Server/Local' badges from the Transcript header for a cleaner UI.
- Added Italian, Japanese, and Hindi to the audio sample clips.
- Fixed greyed-out play buttons by converting OGG samples to real 16 kHz mono PCM WAV.
- Passed explicit language parameter to Whisper generate_kwargs for correct script output.
- Added Cohere language allowlist to gracefully handle unsupported languages.
feat(arch): implement dual-engine ASR testbed and WebGPU baseline
Ping Long
674994f
- Backend: Added HF-GPU/CPU fallback routing to bypass vLLM sm_120 (Blackwell) custom ops crashes.
- Frontend: Integrated Xenova/whisper-base via transformers.js WebGPU client-side inference.
- Frontend: Implemented Web Worker Singleton pattern to cache models and prevent UI spin/OOM.
- Frontend: Added SampleClips UX replicating Hugging Face Spaces (English, Chinese, Spanish).
- Tests: Partitioned Playwright E2E strategy (Client-Side English vs Server-Side Multilingual).
- Scripts: Added fetch_fleurs tool to download native validation audio from Hugging Face APIs.
