config/inference/AV_MossFormer2_TSE_16K.yaml

#!/bin/bash 
mode: 'inference'
use_cuda: 1 # 1 for True, 0 for False
num_gpu: 1
sampling_rate: 16000
network: "AV_MossFormer2_TSE_16K"  # network type
checkpoint_dir: "checkpoints/AV_MossFormer2_TSE_16K"

input_path: "scp/video_samples.scp"  # an input dir or input scp file
output_dir: "path_to_output_videos_tse" # output dir to store processed audio

# decode parameters
one_time_decode_length: 3 # maximum segment length for one-pass decoding (seconds), longer audio (>5s) will use segmented decoding
decode_window: 3 # one-pass decoding length


# Model-specific settings for target speaker extraction
network_reference:
  cue: lip           
  backbone: resnet18  
  emb_size: 256       
network_audio:
  backbone: mossformer2
  encoder_kernel_size: 16
  encoder_out_nchannels: 512
  encoder_in_nchannels: 1
  
  masknet_numspks: 1
  masknet_chunksize: 250
  masknet_numlayers: 1
  masknet_norm: "ln"
  masknet_useextralinearlayer: False
  masknet_extraskipconnection: True

  intra_numlayers: 24
  intra_nhead: 8
  intra_dffn: 1024
  intra_dropout: 0
  intra_use_positional: True
  intra_norm_before: True