Spatially Selective Deep Non-linear Filters for Real-time Multi-channel Speech Enhancement
This website shows a video recording of the real-time demo that we presented at WASPAA 2023.
Key features of the demo system:
- DNN-based spatially selective non-linear filter
- Purely DNN-driven
- Joint spatial and tempo-spectral processing
- High spatial selectivity
- Interactive steering and automatic tracking
- Look direction of the filter is controlled by the user (localization for visual guidance)
- Adjustments of the look direction every 8 ms
- Automatic tracking of slowly moving sources
- Real-time processing
- Processing in the STFT domain (32 ms window, 8 ms shift, 16 kHz sampling frequency)
- Algorithmic latency of 32 ms