support timbre confusion

models/tts/vits/vits_inference.py

@@ -109,6 +109,7 @@ class VitsInference(TTSInference):
             
         # get phone symbol file
         phone_symbol_file = os.path.join(self.exp_dir, self.cfg.preprocess.symbols_dict)
+        print("phone_symbol_file:",phone_symbol_file)
         assert os.path.exists(phone_symbol_file)
         # convert text to phone sequence    
         phone_extractor = phoneExtractor(self.cfg)
@@ -129,26 +130,56 @@ class VitsInference(TTSInference):
             spk2id_file = os.path.join(self.exp_dir, self.cfg.preprocess.spk2id)
             with open(spk2id_file, 'r') as f:
                 spk2id = json.load(f)
-                speaker_id = spk2id[self.args.speaker_name]
-                print("speaker name:",self.args.speaker_name)
-                print("speaker id:",speaker_id)
-                speaker_id = torch.from_numpy(
-                    np.array([speaker_id], dtype=np.int32)
-                ).unsqueeze(0)
+                speaker_name_1= self.args.speaker_name_1
+                speaker_id_1= spk2id[speaker_name_1]
+                print("speaker 1:",speaker_name_1)
+                speaker_id_1 = torch.from_numpy(np.array([speaker_id_1],dtype=np.int32)).unsqueeze(0)
+                    
+                if self.args.speaker_name_2 != (None or "None"):
+                    speaker_name_2= self.args.speaker_name_2
+                    speaker_id_2 = spk2id[speaker_name_2]
+                    print("speaker 2:",speaker_name_2)
+                    speaker_id_2 = torch.from_numpy(np.array([speaker_id_2],dtype=np.int32)).unsqueeze(0)
+                else:
+                    speaker_id_2 = None
                 
         with torch.no_grad():
             x_tst = phone_id_seq.to(self.device).unsqueeze(0)
             x_tst_lengths = torch.LongTensor([phone_id_seq.size(0)]).to(self.device)
-            speaker_id = speaker_id.to(self.device)
-            outputs = self.model.infer(
-                x_tst,
-                x_tst_lengths,
-                sid=speaker_id,
-                noise_scale=noise_scale,
-                noise_scale_w=noise_scale_w,
-                length_scale=length_scale,
-            )
+            if speaker_id_1 is not None:
+                speaker_id_1 = speaker_id_1.to(self.device)
+            if speaker_id_2 is not None:
+                speaker_id_2 = speaker_id_2.to(self.device)
+                outputs_list = []
+                infer_list = [0,1,self.args.alpha]
+                for i in range(3):
+                    outputs_list.append(self.model.infer(
+                        x_tst,
+                        x_tst_lengths,
+                        sid_1=speaker_id_1,
+                        sid_2=speaker_id_2,
+                        alpha=infer_list[i],
+                        noise_scale=noise_scale,
+                        noise_scale_w=noise_scale_w,
+                        length_scale=length_scale,
+                    ))
+                audio_s1=outputs_list[0]["y_hat"][0, 0].data.cpu().float().numpy()
+                audio_s2=outputs_list[1]["y_hat"][0, 0].data.cpu().float().numpy()
+                audio_interpolated=outputs_list[2]["y_hat"][0, 0].data.cpu().float().numpy()
+            else:
+                outputs = self.model.infer(
+                    x_tst,
+                    x_tst_lengths,
+                    sid_1=speaker_id_1,
+                    sid_2=speaker_id_2,
+                    alpha=self.args.alpha,
+                    noise_scale=noise_scale,
+                    noise_scale_w=noise_scale_w,
+                    length_scale=length_scale,
+                )
 
-            audio = outputs["y_hat"][0, 0].data.cpu().float().numpy()
+                audio_s1 = outputs["y_hat"][0, 0].data.cpu().float().numpy()
+                audio_s2 = None
+                audio_interpolated = None
 
-        return audio
+        return audio_s1, audio_s2, audio_interpolated