diff --git a/src/separation/convtasnet_wrapper.py b/src/separation/convtasnet_wrapper.py
new file mode 100644
index 0000000..a1d218e
--- /dev/null
+++ b/src/separation/convtasnet_wrapper.py
@@ -0,0 +1,36 @@
+import torch
+from asteroid.models import ConvTasNet
+
+def separate(audio, sr, model_name='mpariente/ConvTasNet_WHAMR_enhsingle'):
+    """
+    Separates audio into sources using a pretrained Asteroid model.
+
+    Args:
+        audio (numpy.ndarray): The audio time series (1D numpy array).
+        sr (int): Sampling rate of the audio.
+        model_name (str): Name of the pretrained model from Asteroid. Default is 'mpariente/ConvTasNet_WHAMR_enhsingle'.
+
+    Returns:
+        list: List of separated sources as numpy arrays.
+    """
+    try:
+        # Select the device: GPU if available, otherwise CPU
+        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        print(f"Using device: {device}")
+        
+        # Load the pretrained model and move it to the selected device
+        model = ConvTasNet.from_pretrained(model_name).to(device)
+        
+        # Convert the audio array to a PyTorch tensor, add batch dimension, and move to device
+        audio_tensor = torch.tensor(audio).unsqueeze(0).to(device)  # Shape: (1, num_samples)
+        
+        # Perform source separation
+        with torch.no_grad():
+            separated_sources = model(audio_tensor)
+        
+        # Convert separated sources to NumPy arrays and remove batch dimension
+        separated_sources_np = [src.squeeze(0).cpu().numpy() for src in separated_sources]
+        
+        return separated_sources_np
+    except Exception as e:
+        raise RuntimeError(f"Error during separation: {e}")