/se-scaling

Model configurations for scaling SE models in the paper "Beyond Performance Plateaus: A Comprehensive Study on Scalability in Speech Enhancement"

Primary LanguagePython

Description

This repository contains configurations of the following SE models with different model sizes used in the paper Beyond Performance Plateaus: A Comprehensive Study on Scalability in Speech Enhancement:

  • Conv-TasNet
  • DEMUCS-v4
  • BSRNN
  • TF-GridNet

The configurations are based on the ESPnet toolkit and our implementations for the above models are provided in ./espnet2/ as a reference.

Model Causal #Params (M) #MACs (G/s) Config file Model link
16 kHz 48 kHz
BSRNN (sampling-frequency-independent)
xtiny ✔︎ 0.5 0.1 0.4 conf/bsrnn_xtiny.yaml
0.5 0.2 0.6 conf/bsrnn_xtiny_noncausal.yaml Model on HF
tiny ✔︎ 1.3 0.6 1.7 conf/bsrnn_tiny.yaml
1.5 0.7 2.2 conf/bsrnn_tiny_noncausal.yaml Model on HF
small ✔︎ 4.1 2.1 6.4 conf/bsrnn_small.yaml
4.8 2.8 8.5 conf/bsrnn_small_noncausal.yaml Model on HF
medium ✔︎ 14.3 8.4 25.2 conf/bsrnn_medium.yaml
16.9 11.2 33.4 conf/bsrnn_medium_noncausal.yaml Model on HF
large ✔︎ 52.9 33.4 99.9 conf/bsrnn_large.yaml
63.1 44.3 132.5 conf/bsrnn_large_noncausal.yaml Model on HF
xlarge ✔︎ 83.6 66.1 197.7 conf/bsrnn_large_double.yaml
104.1 87.9 262.3 conf/bsrnn_large_double_noncausal.yaml Model on HF

Model Causal #Params (M) #MACs (G/s) Config file Model link
16 kHz 48 kHz
Conv-TasNet (input is always resampled to 48 kHz)
small 1.1 - 8.9 conf/conv_tasnet_small.yaml Model on HF
medium 14.3 - 18.7 conf/conv_tasnet_medium.yaml Model on HF
large 52.6 - 47.2 conf/conv_tasnet_large.yaml Model on HF
xlarge 103.9 - 85.4 conf/conv_tasnet_xlarge.yaml
DEMUCS-v4 (input is always resampld to 48 kHz)
tiny 1.0 - 1.0 conf/demucsv4_tiny.yaml
small 4.1 - 3.5 conf/demucsv4_small.yaml
medium 16.2 - 13.0 conf/demucsv4_medium.yaml
large 26.9 - 17.2 conf/demucsv4_large.yaml
xlarge 79.3 - 40.7 conf/demucsv4_xlarge.yaml
TF-GridNet (sampling-frequency-independent)
xxtiny 0.1 1.9 5.6 conf/tfgridnet_xxtiny.yaml Model on HF
xtiny 0.5 7.4 21.7 conf/tfgridnet_xtiny.yaml Model on HF
tiny 1.5 24.1 70.5 conf/tfgridnet_tiny.yaml Model on HF
small 5.7 89.5 261.8 conf/tfgridnet_small.yaml Model on HF