This PR addresses several critical bugs and feature requests identified in recent sessions, ranging from model loading logic to hardware support and UI/UX improvements.

Key Fixes Applied

Round 1 — Infrastructure & Compatibility

• Issue [Bug] Web version missing @tailwindcss/vite plugin - CSS not loading #121 — CSS Missing in Web Version: Added Tailwind CSS Vite plugin to web/vite.config.ts and updated dependencies.
• Issue not able to find my gpu #127 — "Cannot copy out of meta tensor": Fixed CPU fallback on Apple Silicon by bypassing meta-tensor initialization when using CPU.
• Issue How can I change default download path #125 — Custom Model Directory: Implemented VOICEBOX_MODELS_DIR environment variable support in backend/config.py.

Round 2 — Recording & Upload Stability (Issue #108)

• Bug A (MP3 Uploads): Fixed backend crash when uploading non-WAV files by correctly extracting file extensions.
• Bug B (Duration Mismatch): Replaced fragile duration reading with AudioContext.decodeAudioData for accurate MP3 length detection.
• Bug C (Conversion): Ensured all manually uploaded samples are converted to WAV on the client side before submission.
• Bug D (Recording Race Condition): Fixed a bug where cancelling a recording would still trigger a "complete" event with an empty blob.

Round 3 — Performance & Hardware (Issues #96, #119, #111)

• Issue Bug: Selecting Qwen3-TTS 0.6B still downloads and uses the 1.7B model #96 — 0.6B Model Selection: Fixed race condition where the 1.7B default model was loaded during voice prompt creation before the 0.6B model was initialized.
• Issue Not Support Interl Arc GPU #119 — Intel Arc / DirectML Support: Added hardware detection for Intel XPU (IPEX) and DirectML (torch-directml) backends.
• Issue Very slow for long texts, streaming would be great. #111 — Audio Streaming: Added a new /generate/stream endpoint for instant playback without intermediate disk writes.