Add real-time processing for FRCRN_SE_16K

Related to modelscope#23

Add real-time processing support for the FRCRN_SE_16K model.

* **clearvoice/models/frcrn_se/frcrn.py**
  - Add a new method `real_time_process` to the `FRCRN_SE_16K` class for real-time processing.
  - Modify the `forward` method to support both offline and real-time processing.
  - Update the `DCCRN` class to handle real-time processing.

* **clearvoice/config/inference/FRCRN_SE_16K.yaml**
  - Change `win_len` to 320 to use 20 ms input windows.
  - Change `win_inc` to 160 to use 20 ms input windows.

* **clearvoice/demo.py**
  - Add a new demo case for real-time processing using the `FRCRN_SE_16K` model.

* **clearvoice/demo_with_more_comments.py**
  - Add a new demo case for real-time processing using the `FRCRN_SE_16K` model.

clearvoice/config/inference/FRCRN_SE_16K.yaml

@@ -15,6 +15,6 @@ decode_window: 1 #one-pass decoding length
 #
 # FFT parameters
 win_type: 'hanning'
-win_len: 640
-win_inc: 320
+win_len: 320
+win_inc: 160
 fft_len: 640

clearvoice/demo.py

@@ -65,3 +65,17 @@
 
     #2nd calling method: process video files listed in .scp file, and write outputs to 'path_to_output_videos_tse_scp/'
     myClearVoice(input_path='samples/scp/video_samples.scp', online_write=True, output_path='samples/path_to_output_videos_tse_scp')
+
+##-----Demo Six: use FRCRN_SE_16K model for real-time processing -----------------
+if False:
+    myClearVoice = ClearVoice(task='speech_enhancement', model_names=['FRCRN_SE_16K'])
+
+    ##1st calling method: process an input waveform in real-time and return output waveform, then write to output_FRCRN_SE_16K_realtime.wav
+    output_wav = myClearVoice(input_path='samples/input_realtime.wav', online_write=False)
+    myClearVoice.write(output_wav, output_path='samples/output_FRCRN_SE_16K_realtime.wav')
+
+    ##2nd calling method: process all wav files in 'path_to_input_wavs_realtime/' in real-time and write outputs to 'path_to_output_wavs_realtime'
+    myClearVoice(input_path='samples/path_to_input_wavs_realtime', online_write=True, output_path='samples/path_to_output_wavs_realtime')
+
+    ##3rd calling method: process wav files listed in .scp file in real-time, and write outputs to 'path_to_output_wavs_realtime_scp/'
+    myClearVoice(input_path='samples/scp/audio_samples_realtime.scp', online_write=True, output_path='samples/path_to_output_wavs_realtime_scp')

clearvoice/demo_with_more_comments.py

@@ -57,3 +57,31 @@
         # - online_write (bool): Set to True to enable saving the enhanced output during processing
         # - output_path (str): Path to the directory to save the enhanced output files
         myClearVoice(input_path='samples/scp/audio_samples.scp', online_write=True, output_path='samples/path_to_output_wavs_scp')
+
+    ## ---------------- Demo Three: Real-Time Processing -----------------------
+    if False:  # This block demonstrates how to use the FRCRN_SE_16K model for real-time speech enhancement
+        # Initialize ClearVoice for the task of speech enhancement using the FRCRN_SE_16K model
+        myClearVoice = ClearVoice(task='speech_enhancement', model_names=['FRCRN_SE_16K'])
+
+        # 1st calling method: 
+        #   Process an input waveform in real-time and return the enhanced output waveform
+        # - input_path (str): Path to the input noisy audio file (input_realtime.wav)
+        # - output_wav (dict or ndarray) : The enhanced output waveform
+        output_wav = myClearVoice(input_path='samples/input_realtime.wav', online_write=False)
+        # Write the processed waveform to an output file
+        # - output_path (str): Path to save the enhanced audio file (output_FRCRN_SE_16K_realtime.wav)
+        myClearVoice.write(output_wav, output_path='samples/output_FRCRN_SE_16K_realtime.wav')
+
+        # 2nd calling method: 
+        #   Process and write audio files directly in real-time
+        # - input_path (str): Path to the directory of input noisy audio files
+        # - online_write (bool): Set to True to enable saving the enhanced audio directly to files during processing
+        # - output_path (str): Path to the directory to save the enhanced output files
+        myClearVoice(input_path='samples/path_to_input_wavs_realtime', online_write=True, output_path='samples/path_to_output_wavs_realtime')
+
+        # 3rd calling method: 
+        #   Use an .scp file to specify input audio paths for real-time processing
+        # - input_path (str): Path to a .scp file listing multiple audio file paths
+        # - online_write (bool): Set to True to enable saving the enhanced audio directly to files during processing
+        # - output_path (str): Path to the directory to save the enhanced output files
+        myClearVoice(input_path='samples/scp/audio_samples_realtime.scp', online_write=True, output_path='samples/path_to_output_wavs_realtime_scp')

clearvoice/models/frcrn_se/frcrn.py

@@ -71,18 +71,35 @@ def __init__(self, args):
             win_type=args.win_type
         )
 
-    def forward(self, x):
+    def forward(self, x, real_time=False):
         """
         Forward pass of the model.
 
         Args:
             x (torch.Tensor): Input tensor representing audio signals.
+            real_time (bool): Flag to indicate real-time processing.
 
         Returns:
             torch.Tensor: Processed output tensor after applying the model.
         """
-        output = self.model(x)
-        return output[1][0]  # Return estimated waveform
+        if real_time:
+            return self.real_time_process(x)
+        else:
+            output = self.model(x)
+            return output[1][0]  # Return estimated waveform
+
+    def real_time_process(self, x):
+        """
+        Real-time processing method for the FRCRN model.
+
+        Args:
+            x (torch.Tensor): Input tensor representing audio signals.
+
+        Returns:
+            torch.Tensor: Processed output tensor after applying the model in real-time.
+        """
+        output = self.model.real_time_process(x)
+        return output
 
 
 class DCCRN(nn.Module):
@@ -249,3 +266,41 @@ def get_params(self, weight_decay=0.0):
         }]
         return params
 
+    def real_time_process(self, inputs):
+        """
+        Real-time processing method for the DCCRN model.
+
+        Args:
+            inputs (torch.Tensor): Input tensor representing audio signals.
+
+        Returns:
+            torch.Tensor: Processed output tensor after applying the model in real-time.
+        """
+        out_list = []
+        # Compute the complex spectrogram using STFT
+        cmp_spec = self.stft(inputs)  # [B, D*2, T]
+        cmp_spec = torch.unsqueeze(cmp_spec, 1)  # [B, 1, D*2, T]
+
+        # Split into real and imaginary parts
+        cmp_spec = torch.cat([
+            cmp_spec[:, :, :self.feat_dim, :],  # Real part
+            cmp_spec[:, :, self.feat_dim:, :],  # Imaginary part
+        ], 1)  # [B, 2, D, T]
+
+        cmp_spec = torch.unsqueeze(cmp_spec, 4)  # [B, 2, D, T, 1]
+        cmp_spec = torch.transpose(cmp_spec, 1, 4)  # [B, 1, D, T, 2]
+
+        # Pass through the UNet to estimate masks
+        unet1_out = self.unet(cmp_spec)  # First UNet output
+        cmp_mask1 = torch.tanh(unet1_out)  # First mask
+
+        unet2_out = self.unet2(unet1_out)  # Second UNet output
+        cmp_mask2 = torch.tanh(unet2_out)  # Second mask
+        cmp_mask2 = cmp_mask2 + cmp_mask1  # Combine masks
+
+        # Apply the estimated mask to the complex spectrogram
+        est_spec, est_wav, est_mask = self.apply_mask(cmp_spec, cmp_mask2)
+        out_list.append(est_spec)
+        out_list.append(est_wav)
+        out_list.append(est_mask)
+        return out_list