TABLE IV.
ESTOI (%) scores for different two-stage and single-stage DFNs, LSTMs and BLSTMs in simulated reverberant conditions. In each two-stage DNN, * indicates that the score is significantly better than the masking-based single-stage DNN baseline score (with the significance level of p < 0.0005).
| T60 (s) | 0.3 | 0.6 | 0.9 | |||||
|---|---|---|---|---|---|---|---|---|
| TIR (dB) | −12 | −6 | −12 | −6 | −12 | −6 | Average | |
| Unprocessed | 21.7 | 33.3 | 14.2 | 23.5 | 10.0 | 17.8 | 20.1 | |
| DFN | Single-stage Mapping | 41.9 | 54.9 | 32.3 | 46.4 | 25.5 | 38.5 | 39.9 |
| Single-stage Masking | 49.5 | 62.8 | 37.6 | 51.5 | 29.6 | 43.1 | 45.7 | |
| Mapping+Mapping | 44.8 | 57.7 | 34.1 | 47.6 | 26.6 | 39.6 | 41.7 | |
| Mapping+Masking | 49.4 | 62.7 | 37.1 | 50.9 | 29.0 | 42.2 | 45.2 | |
| Masking+Mapping | 49.6 | 62.0 | 37.5 | 51.2 | 29.2 | 42.9 | 45.4 | |
| Masking+Masking | 53.1* | 66.2* | 40.0* | 53.9* | 31.3* | 45.0* | 48.2* | |
| LSTM | Single-stage Mapping | 48.2 | 60.0 | 37.2 | 50.5 | 29.3 | 42.5 | 44.6 |
| Single-stage Masking | 54.1 | 66.3 | 41.5 | 55.2 | 32.8 | 46.7 | 49.4 | |
| Mapping+Mapping | 49.5 | 60.8 | 37.5 | 50.6 | 29.2 | 42.4 | 45.0 | |
| Mapping+Masking | 53.1 | 65.1 | 40.4 | 53.6 | 31.9 | 45.1 | 48.2 | |
| Masking+Mapping | 53.2 | 64.4 | 40.2 | 53.2 | 31.4 | 44.6 | 47.8 | |
| Masking+Masking | 55.5* | 67.6* | 42.2* | 55.6* | 33.3* | 46.8* | 50.2* | |
| BLSTM | Single-stage Mapping | 52.9 | 64.5 | 41.6 | 54.8 | 32.8 | 46.8 | 48.9 |
| Single-stage Masking | 56.1 | 68.0 | 44.9 | 58.4 | 35.4 | 50.1 | 52.1 | |
| Mapping+Mapping | 53.7 | 64.3 | 40.5 | 54.6 | 30.4 | 45.4 | 48.1 | |
| Mapping+Masking | 54.2 | 65.4 | 41.8 | 55.4 | 32.2 | 46.9 | 49.3 | |
| Masking+Mapping | 57.4* | 68.6* | 44.8 | 58.5 | 34.5 | 49.4 | 52.2 | |
| Masking+Masking | 58.0* | 69.8* | 45.5* | 59.3* | 36.0* | 50.8* | 53.2* | |