fix: improve format conversion robustness and align tensor shapes for dry/wet blending

app.py

@@ -104,10 +104,18 @@ def load_audio_gradio(
 
 
 def ensure_wav(filepath: str) -> str:
-    """Convert MP3 (or other formats) to WAV using ffmpeg if needed."""
-    if filepath.lower().endswith(".mp3"):
+    """Convert MP3 (or other non-WAV formats) to WAV using ffmpeg if needed."""
+    if not filepath.lower().endswith(".wav"):
         wav_path = filepath.rsplit(".", 1)[0] + ".wav"
-        subprocess.run(["ffmpeg", "-y", "-i", filepath, wav_path], check=True)
+        if wav_path == filepath:
+            wav_path = filepath + ".wav"
+        # Run ffmpeg silently to convert the format
+        subprocess.run(
+            ["ffmpeg", "-y", "-i", filepath, wav_path],
+            check=True,
+            stdout=subprocess.DEVNULL,
+            stderr=subprocess.DEVNULL
+        )
         return wav_path
     return filepath
 
@@ -161,6 +169,11 @@ def demo_fn(
     logger.info("Denoising finished")
 
     # Dry/wet blending: mix original 'sample' back into 'enhanced' to restore voice texture
+    # Ensure both tensors have the exact same shape (DeepFilterNet can output a slightly different length due to STFT framing)
+    min_len = min(enhanced.shape[-1], sample.shape[-1])
+    enhanced = enhanced[..., :min_len]
+    sample = sample[..., :min_len]
+
     alpha = wet_dry_mix / 100.0
     enhanced = alpha * enhanced + (1 - alpha) * sample