noise-suppression-denseunet-ll-0001

Use Case and High-Level Description

This is a model for noise suppression to make speech cleaner. The model architecture is similar to PoCoNet, Channel-Attention Dense U-Net but without multi head attensions (MHA) to decrease model complexity and increse processing speed. Also to reduce processed patch size and achieve small delay the pyramid structure along time axis is reduced and convolutions with dilation along time axis are used. The model was trained on DNS-Challenge dataset paper. The network works with mono audio sampled on 16kHz. The audio is processed iteratively by patches with 128 size. On each iteration it takes 128 (8ms) samples as input and returns 128 (8ms) samples as output with 384 (24ms) samples delay. In addition the network required 39 state tensors to make processing seamless. On the first iteration these state tensors have to be filled with 0. On the consequences iterations these tensors have to be taken from corresponding outputs of previous iteration. You can try Noise Suppression CPP* Demo to see how it works.

Specification

to process 128 samples that is 8ms for 16kHz

Metric

Value

GOps

0.2

MParams

4.2

Source framework

PyTorch*

Accuracy

The SISDR quality metric was calculated on the 100 dev test synthetic speech clips from DNS-Challenge dataset.

Metric

Value

SISDR for input noisy signal

11.7 dB

SISDR for output cleaned signal

20.0 dB

SISDR increase

+8.3 dB

Input

Sequence patch, name: input, shape: 1, 128, format: B, T, where:

  • B - batch size

  • T - number of samples in patch

input states, names: inp_state_*, should be filled by corresponding out_state_* from previous step. Total number of input states is 39

Output

Sequence patch, name: output, shape: 1, 128, format: B, T, where:

  • B - batch size

  • T - number of samples in patch

Note: The output patch is “shifted” by 384 (24ms) samples in time. So output[0,i] sample is synced with input[0,i-384] sample

output states, names: out_state_*, should be used to fill corresponding inp_state_* on next step. Total number of output states is 39

The next outputs are optional to process

Output spectrum, name: Y, shape: 1, 2, 129, 1, format: B, C, F, T, where:

  • B - batch size

  • C - number of complex value components (always 2)

  • F - number of frequency bins

  • T - number of spectrums in patch

Output delay, name: delay, shape: 1, format: C, where:

  • C - number of delays (always 1) This is provided for convenience and contains the output “shift” relative to the input. For this model it is 384 and can differ for other noise suppression models.

Demo usage

The model can be used in the following demos provided by the Open Model Zoo to show its capabilities: