Ensemble detector (--detector ensemble)¶

The recommended detector. Combines multiple detectors and uses a voting algorithm to pick the most confident classification for each segment.

Usage¶

praisonai-editor edit file.mp3 --preset songs_only --detector ensemble

The --detector auto option also routes to ensemble.

How it works¶

flowchart LR
    A[Audio] --> B[FFmpeg astats]
    A --> C[Librosa spectral]
    A --> D[INA CNN\nif installed]
    A --> E[Demucs stems\nif --demix]
    B --> F[Ensemble decision\nconflict resolution]
    C --> F
    D --> F
    E --> F
    F --> G[Classified blocks\nwith confidence]

Conflict resolution rules: - If INA is available → INA has highest priority for speech/music boundaries - If Demucs is used → demix scores refine singing vs talking-over-music - FFmpeg provides base signal statistics for all segments

Confidence scores¶

Each block gets a confidence score (0–1). Low-confidence borders are re-classified using weighted voting from all available detectors.