In order to verify the correctness and speed of the implementations in Julius, we compare ourselves to different reference implementations, comparing speed and checking how far we are.
We compare julius.resample to resampy, on an input of size (32, 8 * 44100),
i.e. a batch of size 16 of 8 second of audio at 44.1kHz.
We use the same number of zero crossing as resampy for this benchmark.
The small delta is probably
due to the different window function used.
On CPU we have:
| Old sr | New sr | Julius (ms) | Resampy (ms) | Delta (%) |
|---|---|---|---|---|
| 2 | 1 | 310 | 1639 | 1.3% |
| 1 | 2 | 380 | 2646 | 1.8% |
| 4 | 5 | 94 | 1704 | 1.8% |
| 10 | 11 | 53 | 1479 | 1.8% |
| 44100 | 16000 | 70 | 2552 | 0.9% |
| 20001 | 30001 | 23788 | 2546 | 1.8% |
On GPU we have:
| Old sr | New sr | Julius (ms) |
|---|---|---|
| 2 | 1 | 44 |
| 1 | 2 | 8 |
| 4 | 5 | 5 |
| 10 | 11 | 6 |
| 44100 | 16000 | 23 |
| 20001 | 30001 | 15186 |
We compare to pytorch.nn.functional.conv1d, on a input of size [32, 32, 10240],
for a convolution with 32 input channels, 64 output channels and various kernel sizes.
On CPU we have:
| Kernel size | FFT (ms) | No FFT (ms) | Delta |
|---|---|---|---|
| 8 | 506 | 204 | 4.0e-06 |
| 32 | 410 | 409 | 1.4e-05 |
| 64 | 400 | 847 | 2.7e-05 |
| 128 | 356 | 1854 | 5.3e-05 |
| 256 | 350 | 3315 | 1.0e-04 |
| 1024 | 683 | 17779 | 4.1e-04 |
| 2048 | 1031 | 28133 | 8.2e-04 |
On GPU we have:
| Kernel size | FFT (ms) | No FFT (ms) | Delta |
|---|---|---|---|
| 8 | 15 | 4 | 4.4e-06 |
| 32 | 12 | 5 | 1.5e-05 |
| 64 | 10 | 11 | 3.1e-05 |
| 128 | 10 | 22 | 5.4e-05 |
| 256 | 9 | 44 | 1.1e-04 |
| 1024 | 14 | 330 | 4.2e-04 |
| 2048 | 13 | 561 | 8.3e-04 |
We do not compare to anything, but measure the attenuation in dB of a pure tone
at 0.9 * cutoff, at the cutoff, and at 1.1 * cutoff.
Note that our implementation automatically choses to use FFTConv1d or not when appropriate.
On CPU we have:
| Freq. | Attn. 0.9 (dB) | Attn 1.0 (dB) | Attn 1.1 (dB) | Time (ms) |
|---|---|---|---|---|
| 0.005 | -1.41 | -6.02 | -16.41 | 9 |
| 0.01 | -1.41 | -6.02 | -16.46 | 7 |
| 0.1 | -1.41 | -6.02 | -16.48 | 11 |
| 0.2 | -1.41 | -6.02 | -16.48 | 5 |
| 0.4 | -1.41 | -6.03 | -16.38 | 6 |
On GPU we have:
| Freq. | Attn. 0.9 (dB) | Attn 1.0 (dB) | Attn 1.1 (dB) | Time (ms) |
|---|---|---|---|---|
| 0.005 | -1.41 | -6.02 | -16.41 | 1 |
| 0.01 | -1.41 | -6.02 | -16.46 | 1 |
| 0.1 | -1.41 | -6.02 | -16.48 | 1 |
| 0.2 | -1.41 | -6.02 | -16.48 | 0 |
| 0.4 | -1.41 | -6.03 | -16.38 | 0 |