update

examples/fsmn_vad_by_webrtcvad/run.sh

@@ -7,7 +7,7 @@ bash run.sh --stage 3 --stop_stage 5 --system_version centos \
 --final_model_name fsmn-vad-nx2-dns3-256-128-4-20251125 \
 --noise_patterns "/data/tianxing/HuggingDatasets/nx_noise/data/noise/**/*.wav" \
 --speech_patterns "/data/tianxing/HuggingDatasets/nx_noise/data/speech/dns3-speech/**/*.wav \
-/data/tianxing/HuggingDatasets/nx_noise/data/speech/nx-speech2/**/*.wav"
+/data/tianxing/HuggingDatasets/nx_noise/data/speech/nx-speech2/**/*.wav" \
 --config_file yaml/config-256-128-4.yaml
 
 

examples/fsmn_vad_by_webrtcvad/yaml/config-240-80-4.yaml

@@ -2,7 +2,7 @@ model_name: "fsmn_vad"
 
 # spec
 sample_rate: 8000
-nfft: 512
+nfft: 256
 win_size: 240
 hop_size: 80
 win_type: hann

examples/fsmn_vad_by_webrtcvad/yaml/config-256-128-4.yaml

@@ -2,7 +2,7 @@ model_name: "fsmn_vad"
 
 # spec
 sample_rate: 8000
-nfft: 512
+nfft: 256
 win_size: 256
 hop_size: 128
 win_type: hann

examples/fsmn_vad_by_webrtcvad/yaml/config-512-256-4.yaml

@@ -0,0 +1,46 @@
+model_name: "fsmn_vad"
+
+# spec
+sample_rate: 8000
+nfft: 512
+win_size: 256
+hop_size: 128
+win_type: hann
+
+# model
+fsmn_input_size: 513
+fsmn_input_affine_size: 140
+fsmn_hidden_size: 250
+fsmn_basic_block_layers: 4
+fsmn_basic_block_hidden_size: 128
+fsmn_basic_block_lorder: 20
+fsmn_basic_block_rorder: 0
+fsmn_basic_block_lstride: 1
+fsmn_basic_block_rstride: 0
+fsmn_output_affine_size: 140
+fsmn_output_size: 2
+
+# lsnr
+n_frame: 3
+min_local_snr_db: -15
+max_local_snr_db: 30
+norm_tau: 1.
+
+# data
+min_snr_db: -10
+max_snr_db: 20
+
+# train
+lr: 0.001
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
+
+max_epochs: 100
+clip_grad_norm: 10.0
+seed: 1234
+
+num_workers: 4
+batch_size: 128
+eval_steps: 25000

examples/silero_vad_by_webrtcvad/yaml/config-512-256-4.yaml

@@ -0,0 +1,43 @@
+model_name: "silero_vad"
+
+# spec
+sample_rate: 8000
+nfft: 512
+win_size: 512
+hop_size: 256
+win_type: hann
+
+# model
+encoder_in_channels: 64
+encoder_hidden_channels: 128
+encoder_out_channels: 128
+encoder_kernel_size: 3
+encoder_num_layers: 4
+
+decoder_hidden_size: 128
+decoder_num_layers: 2
+
+# lsnr
+n_frame: 3
+min_local_snr_db: -15
+max_local_snr_db: 30
+norm_tau: 1.
+
+# data
+min_snr_db: -10
+max_snr_db: 20
+
+# train
+lr: 0.001
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
+
+max_epochs: 100
+clip_grad_norm: 10.0
+seed: 1234
+
+num_workers: 4
+batch_size: 128
+eval_steps: 25000