Table 1.
Objective measure scores: HIT-FA alarm rates (with FA scores in brackets), mean square error (MSE) between ERM and IRM, and NCM and STOI scores for the RNN algorithms used in the listening experiment, RNN-B and RNN-T, and for UN and IRM in both test noise conditions (20-talker babble and traffic noises) and three SNRs. The table shows results both for matched-noise (RNN-B in babble, RNN-T in traffic) and unmatched-noise (RNN-B in traffic, RNN-T in babble) conditions between training and testing.
| Tested with babble noise | Tested with traffic noise | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Metric | SNR | UN | RNN-B | RNN-T | IRM | UN | RNN-T | RNN-B | IRM |
| HIT-FA (FA) | 0 | 65 (18) | 30 (53) | 74 (18) | 71 (14) | ||||
| 5 | 78 (9) | 46 (42) | 80 (10) | 77 (9) | |||||
| 10 | 82 (3) | 62 (25) | 84 (6) | 79 (4) | |||||
| MSE | 0 | 0.079 | 0.230 | 0.064 | 0.061 | ||||
| 5 | 0.039 | 0.170 | 0.037 | 0.041 | |||||
| 10 | 0.028 | 0.100 | 0.028 | 0.036 | |||||
| STOI | 0 | 0.71 | 0.77 | 0.74 | 0.91 | 0.82 | 0.86 | 0.85 | 0.94 |
| 5 | 0.82 | 0.87 | 0.86 | 0.94 | 0.90 | 0.92 | 0.91 | 0.96 | |
| 10 | 0.90 | 0.93 | 0.92 | 0.96 | 0.94 | 0.95 | 0.95 | 0.98 | |
| NCM | 0 | 0.55 | 0.71 | 0.63 | 0.92 | 0.72 | 0.81 | 0.79 | 0.96 |
| 5 | 0.74 | 0.87 | 0.82 | 0.96 | 0.85 | 0.91 | 0.90 | 0.98 | |
| 10 | 0.89 | 0.95 | 0.94 | 0.99 | 0.93 | 0.96 | 0.95 | 0.99 | |