Update app.py

app.py

@@ -215,8 +215,14 @@ def infer_inp(prompt, audio_path, mask_start_point, mask_end_point, progress=gr.
     # Ensure correct shape
     denorm_spec_audio = denorm_spec_audio.flatten()  # Converts (1, N) → (N,)
 
-    # Normalize the audio to prevent clipping or excessive loudness
-    denorm_spec_audio = denorm_spec_audio / np.max(np.abs(denorm_spec_audio))  # Scale between -1 and 1
+    print("Original spectrogram min/max:", norm_spec.min().item(), norm_spec.max().item())
+    print("Generated spectrogram min/max:", output_spec.min().item(), output_spec.max().item())
+
+    global_max = max(np.max(np.abs(raw_chunk_audio)), np.max(np.abs(denorm_spec_audio)))
+    denorm_spec_audio = denorm_spec_audio / global_max  # Normalize using a shared scale
+
+    print("Masked spectrogram min/max:", masked_spec.min().item(), masked_spec.max().item())
+    print("Output spectrogram min/max:", output_spec.min().item(), output_spec.max().item())
 
     # Save as WAV
     sf.write("output.wav", denorm_spec_audio, 16000)
@@ -394,6 +400,12 @@ with gr.Blocks(css=css) as demo:
                 outputs = [audio_in_spec]
             )
 
+            audio_in_inp.stop_recording(
+                fn = load_input_spectrogram,
+                inputs = [audio_in_inp],
+                outputs = [audio_in_spec]
+            )
+
             preview_mask_btn.click(
                 fn = preview_masked_area,
                 inputs = [audio_in_inp, mask_start_point, mask_end_point],