Use Case and High-Level Description

This is a PoCoNet like model for noise suppression to make speech cleaner. The model is based on PoCoNet architecure and trained on subset of DNS-Challenge dataset paper. The subset consists of data derived from LibriVox, VocalSet, CREMA-D, M-AILABS, AudioSet, Freesound, DEMAND datasets. The network works with mono audio sampled on 16kHz. The audio processed iterative by patches with 2048 size. On each iteration it takes 2048 (128ms) samples as input and returns 2048 (128ms) samples as output with 640 (40ms) samples delay. In addition the network required 52 state tensors to make processing seamless. On the first iteration these state tensors have to be filled with 0. On the consequences iterations theses tensors have to be taken from corresponding outputs of previous iteration. You can try Noise Suppression Python* Demo to see how it works.


to process 2048 samples that is 128ms for 16kHz







Source framework



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



SISDR for input noisy signal

11.73 dB

SISDR for output cleaned signal

19.84 dB

SISDR increase

+8.10 dB


Sequence patch, name: input, shape: 1, 2048, 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


Sequence patch, name: output, shape: 1, 2048, format: B, T, where:

  • B - batch size

  • T - number of samples in patch Note: The output patch is “shifted” by 640 (40ms) samples in time. So output[0,i] sample is synced with input[0,i-640] sample

output states, names: out_state_*, should be used to fill corresponding inp_state_* on next step