update

examples/nx_denoise/run.sh

@@ -3,27 +3,11 @@
 : <<'END'
 
 
-sh run.sh --stage 2 --stop_stage 2 --system_version windows --file_folder_name file_dir --final_model_name mpnet-aishell-20250224 \
---noise_dir "E:/Users/tianx/HuggingDatasets/nx_noise/data/noise" \
---speech_dir "E:/programmer/asr_datasets/aishell/data_aishell/wav/train"
-
-
-sh run.sh --stage 3 --stop_stage 3 --system_version centos --file_folder_name file_dir --final_model_name mpnet-aishell-20250224 \
---noise_dir "/data/tianxing/HuggingDatasets/nx_noise/data/noise" \
---speech_dir "/data/tianxing/HuggingDatasets/aishell/data_aishell/wav/train"
-
-sh run.sh --stage 2 --stop_stage 2 --system_version centos --file_folder_name file_dir --final_model_name nx-clean-unet-aishell-20250228 \
+sh run.sh --stage 2 --stop_stage 2 --system_version centos --file_folder_name file_dir --final_model_name nx-denoise-aishell-20250228 \
 --noise_dir "/data/tianxing/HuggingDatasets/nx_noise/data/noise" \
 --speech_dir "/data/tianxing/HuggingDatasets/aishell/data_aishell/wav/train" \
 --max_epochs 100
 
-
-sh run.sh --stage 1 --stop_stage 2 --system_version centos --file_folder_name file_dir --final_model_name mpnet-nx-speech-20250224 \
---noise_dir "/data/tianxing/HuggingDatasets/nx_noise/data/noise" \
---speech_dir "/data/tianxing/HuggingDatasets/nx_noise/data/speech" \
---max_epochs 100 --max_count 10000
-
-
 END
 
 

examples/nx_denoise/yaml/config.yaml

@@ -5,16 +5,7 @@ segment_size: 16000
 n_fft: 512
 win_size: 200
 hop_size: 80
-# 因为 hop_size 取 80，则相当于 stft 的时间步是 10ms 一步，所以降采样也考虑到差不多的分辨率。
 
-# 2**down_sampling_num_layers，
-# 例如 2**6=64 就意味着 64 个值在降采样之后是一个时间步，
-# 则一步是 64/sample_rate = 0.008秒。
-# 那么 tsfm_chunk_size=2 则为16ms，tsfm_chunk_size=4 则为32ms
-# 假设每次向左看1秒，向右看30ms，则：
-# tsfm_chunk_size=1，tsfm_num_left_chunks=128，tsfm_num_right_chunks=4
-# tsfm_chunk_size=2，tsfm_num_left_chunks=64，tsfm_num_right_chunks=2
-# tsfm_chunk_size=4，tsfm_num_left_chunks=32，tsfm_num_right_chunks=1
 down_sampling_num_layers: 6
 down_sampling_in_channels: 1
 down_sampling_hidden_channels: 64

toolbox/torchaudio/models/nx_denoise/modeling_nx_denoise.py

@@ -244,6 +244,8 @@ class NXDenoise(nn.Module):
 
         # ts transformer
         # bottle_neck shape: [batch_size, channels, time_steps, freq_dim]
+        bottle_neck = self.ts_transformer.forward(bottle_neck)
+        # bottle_neck shape: [batch_size, channels, time_steps, freq_dim]
 
         # causal conv out
         bottle_neck = self.causal_conv_out.forward(bottle_neck)