Optimize: Lower Chunk Threshold to 0.90 to fix AI false negatives (relying on Heuristic Override for false positive protection)

Optimize: Tighten Heuristic thresholds and disable dithering to fix AI false negatives

Optimize: Implement Heuristic Override (Spectral Check) to prevent false positives on compressed audio

Optimize: Robust M4A handling and Noise Injection (Dithering) to prevent compression artifacts false positives

Optimize: Implement silent chunk filtering and majority voting with 99% threshold

Optimize: Implement robust sliding window inference and re-enable normalization

Fix: Disable silence trimming/normalization and optimize pipeline input

Optimize accuracy: Add 95% confidence threshold for AI classification

Deploy optimized VoiceGuard with Abhishtagatya model

Critical Fix: Invert label logic (Label 0 is Fake, Label 1 is Real)

Simplify to robust 2-model ensemble (WavLM + Melody)

Update to 3-model weighted ensemble (HuBERT/WavLM/Melody) and add torch requirements

Upgrade to WavLM + Unified Label Parsing

Optimize for Indian Accents: Silence Trim + Conservative Voting

Fix MelodyMachine label parsing (LABEL_1=AI)

Upgrade to Dual-Model Ensemble (HuBERT + MelodyMachine)

Switch to Hybrid Detector (ML + Signal Analysis)

Fix logic: Remove 98% AI confidence threshold

Deploy to Spaces