Restore the input's sample rate and channel count

app.py

@@ -19,10 +19,15 @@ model_card = ModelCard(
     name="DeepFilterNet",
     description=(
         (
-            "Background noise suppression / speech enhancement using DeepFilterNet3. "
-            "Input is converted to mono 48kHz. Slider controls strength through "
-            "wet/dry blend. Repository: https://github.com/Rikorose/DeepFilterNet "
+            (
+            "Background noise suppression / speech enhancement using "
+            "DeepFilterNet3. Input is converted to mono 48kHz. Slider controls "
+            "strength through wet/dry blend. Denoising runs at 48 kHz in mono; the "
+            "result is resampled back to the input's rate and copied across its "
+            "channels, so the output matches the format it came in with but carries "
+            "no stereo image. Repository: https://github.com/Rikorose/DeepFilterNet "
             "Paper: https://arxiv.org/abs/2305.08227"
+            )
         )
     ),
     author="Hendrik Schröter, Alberto N. Escalante-B., Tobias Rosenkranz, Andreas Maier",
@@ -87,6 +92,12 @@ def process_fn(input_audio_path: str, noise_atten_db: float) -> str:
     if not input_audio_path:
         raise ValueError("No input audio provided.")
 
+    # The session's format, read before the downmix and resample that
+    # DeepFilterNet needs, and restored before saving
+    info = sf.info(input_audio_path)
+    original_sr = info.samplerate
+    original_channels = info.channels
+
     # Load and normalize
     noisy, sr = load_audio_mono_48k(input_audio_path)
 
@@ -105,7 +116,15 @@ def process_fn(input_audio_path: str, noise_atten_db: float) -> str:
     out_dir.mkdir(parents=True, exist_ok=True)
     out_path = out_dir / "denoised.wav"
 
-    sf.write(str(out_path), out, sr)
+    if original_sr != sr:
+        out = librosa.resample(out, orig_sr=sr, target_sr=original_sr)
+
+    # Denoising is mono here, so the channels are copies rather than a
+    # recovered image, which the model card states
+    if original_channels > 1:
+        out = np.tile(out[:, None], (1, original_channels))
+
+    sf.write(str(out_path), out, original_sr)
     return str(out_path)