Direction Specific Ambisonics Source Separation with End-To-End Deep Learning - Listening Examples
This webpage provides listening examples for the paper "Direction Specific Ambisonics Source Separation with End-To-End Deep Learning". First, we provide listening examples corresponding to Figure 4 from the paper. Specifically, we provide the audio predicted by each method at the direction of each source (illustrated as red dots in the figure).Figure 4 Listening Examples (1st order, anechoic, musdb). Mixture:
| GT signal | Refinement | Implicit | Mixed | max-re | |
| Direction Middle Dot | |||||
| Direction Left Dot | |||||
| Direction Right Dot |
| Refinement | Implicit | Mixed | max-re | ||
| Direction Middle Dot | 10.74 dB | 13.36 dB | 15.18 dB | 6.92 dB | |
| Direction Left Dot | 4.94 dB | 9.15 dB | 10.07 dB | 1.77 dB | |
| Direction Right Dot | 10.22 dB | 8.95 dB | 13.11 dB | 10.45 dB |
Below you can find listening examples for a mixture generated with the Musdb18 dataset under anechoic conditions and a mixture generated with the FUSS dataset under room conditions. For each mixture, we provide the predictions of the method considering the Ambisonics mixture encoded using first and fourth order. Each mixture contains three sources at directions 1, 2, and 3. We show the ground truth and predicted signals when using the direction of each source as the target direction for all methods. In addition, we also provide the predictions corresponding to a source-free region, i.e. direction 4 (Azi:0 Zen:0).
Musb18 Anechoic Example (1st order). Mixture:
Scroll sideways to find more samples.
This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 812719.