neural net running at 192kHz to wave shape input quadrature X into 4 different soft clip waves based on additional input 2d embedding
corners here represent the training data; all data has a soft clip applied
- e0,e1 = -1, -1 => sine wave
- e0,e1 = -1, 1 => square wave
- e0,e1 = 1, 1 => zigzag like wave
- e0,e1 = -1, 1 => saw wave
0 core sine wave waveshaping input
1 core cosine wave waveshaping input ( normalled to phase shfted estimate from in0 )
2 embedding x
3 embedding y
0 waveshaped result ( with small amount of low pass filtering )
1 raw waveshaped result
2 -
3 cosine estimated from in0 sine
- support audio rate modulation of all inputs
- trained on core sine wave, but takes any wave as input
- the in1 derived from phase shifted in0 can be noisy under heavy in0 FM
- inference runs at 192kHz
- pretrain at FP 3.15, fine tune at FP 3.5
- receptive field of 4096 samples => ~9 cycles of a 440Hz wave at 192kHz
- pretrain with MSE with ramped up STFT losses; finetune with just MSE
- trained on synthetic sampled data at frequencies between A2 to A6
- receptive field of 1024 samples ( => ~2.3 cycles of a 440Hz wave at 192kHz )
- double width biases and accumulation
- EBR activation caching
Layer (type) Output Shape Param #
==========================================================
input_1 (InputLayer) (None, 4096, 4) 0
qconv_0_qb (QConv1D) (None, 1024, 8) 136
qconv_1_qb (QConv1D) (None, 256, 16) 528
qconv_2_qb (QConv1D) (None, 64, 16) 1040
qconv_3_qb (QConv1D) (None, 16, 8) 520
qconv_4_qb (QConv1D) (None, 4, 8) 264
qconv_5_qb (QConv1D) (None, 1, 1) 33
==========================================================
Total params: 2521 (9.85 KB)
( FP3.5 at 192kHz )
DP16KD:75% MULT18X18D:46% ALU54B:0% TRELLIS_FF:54% TRELLIS_COMB:61%
Info: Router1 time: 0h 01m 06s
Info: '$glbnet$audio_clk': 65.58 MHz (PASS at 50.00 MHz)
Info: '$glbnet$clk': 79.65 MHz (PASS at 60.00 MHz)