How thomeou/salsa-lite works
SALSA-Lite is a research workflow for polyphonic sound event localization and detection (SELD): multichannel microphone audio is converted into SALSA-Lite features, then used to train or run a CRNN that detects sound classes and estimates event directions and timing. The README also describes dataset preparation, alternative feature comparisons, and evaluation results. The supplied tree contains no implementation files, so component paths and internal call wiring cannot be verified; the links below reflect only documented data flow.
Components 12
Dataset
- MIC audio
- Event metadata
- Dataset preparation
Feature extraction
- Feature extraction
- SALSA-Lite NIPD
- Feature alternatives
SELD model
- CRNN training
- CRNN inference
- SELD events
Evaluation
- Performance evaluation
- SELD metrics
Other components
- Researcher
Connections 14
- Researcher → Dataset preparation: prepares data
- MIC audio → Dataset preparation: provides recordings
- Event metadata → Dataset preparation: provides labels
- Dataset preparation → Feature extraction: supplies audio
- Feature extraction → SALSA-Lite NIPD: computes features
- SALSA-Lite NIPD → CRNN training: provides features
- Event metadata → CRNN training: provides targets
- CRNN training → CRNN inference: produces model
- SALSA-Lite NIPD → CRNN inference: provides features
- CRNN inference → SELD events: predicts events
- SELD events → Performance evaluation: supplies predictions
- Event metadata → Performance evaluation: supplies references
- Performance evaluation → SELD metrics: reports metrics
- Researcher → Feature alternatives: compares features