Table 3. Comparison to alternative methods.
Methods used for comparisons: (1) Arthur et al., 2013, (2) Tachibana et al., 2020, (3) Oikarinen et al., 2019, (4) Cohen et al., 2020. (A,B) DAS was trained on 1825/15970 syllables which contained 4/7 of the call types from Oikarinen et al., 2019. (B) The method by Oikarinen et al., 2019 produces an annotation every 50 ms of the recording - since the on/offset can occur anywhere within the 50 ms, the expected error of the method by Oikarinen et al., 2019 is at least 12.5 ms. (C) The method by Oikarinen et al., 2019 annotates 60 minutes of recordings in 8 minutes. (D) Throughput assessed on the CPU, since the methods by Arthur et al., 2013 and Tachibana et al., 2020 do not run on a GPU. (E) Throughput assessed on the GPU. The methods by Cohen et al., 2020 and Oikarinen et al., 2019 use a GPU.
| Precision [%] | Recall [%] | Jitter [ms] | Throughput [s/s] | |||||
|---|---|---|---|---|---|---|---|---|
| Species | DAS | Other | DAS | Other | DAS | Other | DAS | Other |
| Fly single (1) | 97/92 (p/s) | 99/91 | 96/98 (p/s) | 87/91 | 0.3/12 (p/s) | 0.1/22 | 15 | 4 (D) |
| Fly multi (1) | 98 | 99 | 94 | 92 | 0.3 | 0.1 | 8 (p) | 0.4 (p+s) (D) |
| Fly multi (1) | 97 | 95 | 93 | 93 | 8.0 | 15.0 | 8 (s) | 0.4 (p+s) (D) |
| Mouse (2) | 98 | 98 | 99 | 99 | 0.3 | 0.4 | 12 | 4 (D) |
| Marmoset (3) | 96 | 85 (A) | 92 | 77 (A) | 4.4 | 12.5 (B) | 82 | 7.5 (C, E) |
| Bengalese finch (4) | 99 | 99 | 99 | 99 | 0.3 | 1.1 | 15 | 5 (E) |
| Zebra finch (4) | 100 | 100 | 100 | 100 | 1.3 | 2.0 | 18 | 5 (E) |