Update inference/style_transfer.py

inference/style_transfer.py

@@ -22,6 +22,7 @@ sys.path.append(os.path.join(os.path.dirname(currentdir), "mixing_style_transfer
 from networks import FXencoder, TCNModel
 from data_loader import *
 import librosa
+import pyloudnorm
 
 
 
@@ -170,8 +171,20 @@ class Mixing_Style_Transfer_Inference:
                 # save output of each instrument
                 if self.args.save_each_inst:
                     sf.write(os.path.join(cur_out_dir, f"{cur_inst_name}_{output_name_tag}.wav"), fin_data_out_inst.transpose(-1, -2), self.args.sample_rate, 'PCM_16')
+                    
             # remix
             fin_data_out_mix = sum(inst_outputs)
+            
+            # loudness adjusting for mastering purpose
+            meter = pyloudnorm.Meter(44100)
+            loudness_out = meter.integrated_loudness(fin_data_out_mix.transpose(-1, -2))
+            reference_aud = load_wav_segment(reference_track_path, axis=1)
+            loudness_ref = meter.integrated_loudness(reference_aud)
+            # adjust output loudness to that of the reference
+            fin_data_out_mix = pyloudnorm.normalize.loudness(fin_data_out_mix, loudness_out, loudness_ref)
+            fin_data_out_mix = np.clip(fin_data_out_mix, -1., 1.)
+            
+            # save output
             fin_output_path = os.path.join(cur_out_dir, f"mixture_{output_name_tag}.wav")
             sf.write(fin_output_path, fin_data_out_mix.transpose(-1, -2), self.args.sample_rate, 'PCM_16')