fix voice cloning of vits.

examples/aishell3/vits-vc/README.md

@@ -122,13 +122,13 @@ ref_audio
 `./local/voice_cloning.sh` calls `${BIN_DIR}/voice_cloning.py`
 
 ```bash
-CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} ${ge2e_params_path} ${ref_audio_dir}
+CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} ${ge2e_params_path} ${add_blank} ${ref_audio_dir}
 ```
 
 If you want to convert a speaker audio file to refered speaker, run:
 
 ```bash
-CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} ${ge2e_params_path} ${ref_audio_dir} ${src_audio_path}
+CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} ${ge2e_params_path} ${add_blank} ${ref_audio_dir} ${src_audio_path}
 ```
 
 ## Pretrained Model

examples/aishell3/vits-vc/local/voice_cloning.sh

@@ -4,8 +4,8 @@ config_path=$1
 train_output_path=$2
 ckpt_name=$3
 ge2e_params_path=$4
-ref_audio_dir=$5
-add_blank=$6
+add_blank=$5
+ref_audio_dir=$6
 src_audio_path=$7
 
 FLAGS_allocator_strategy=naive_best_fit \
@@ -14,9 +14,9 @@ python3 ${BIN_DIR}/voice_cloning.py \
     --config=${config_path} \
     --ckpt=${train_output_path}/checkpoints/${ckpt_name} \
     --ge2e_params_path=${ge2e_params_path} \
+    --phones_dict=dump/phone_id_map.txt \
     --text="凯莫瑞安联合体的经济崩溃迫在眉睫。" \
     --audio-path=${src_audio_path} \
     --input-dir=${ref_audio_dir} \
     --output-dir=${train_output_path}/vc_syn \
-    --phones-dict=dump/phone_id_map.txt \
     --add-blank=${add_blank}

examples/aishell3/vits-vc/run.sh

@@ -11,6 +11,7 @@ conf_path=conf/default.yaml
 train_output_path=exp/default
 ckpt_name=snapshot_iter_153.pdz
 add_blank=true
+ref_audio_dir=ref_audio
 src_audio_path=''
 
 # not include ".pdparams" here
@@ -40,5 +41,5 @@ fi
 
 if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
     CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} \
-        ${ge2e_params_path} ${ref_audio_dir} ${add_blank} ${src_audio_path} || exit -1
+        ${ge2e_params_path} ${add_blank} ${ref_audio_dir} ${src_audio_path} || exit -1
 fi

paddlespeech/t2s/exps/vits/synthesize.py

@@ -52,9 +52,8 @@ def evaluate(args):
         spk_num = len(spk_id)
         fields += ["spk_id"]
     elif args.voice_cloning:
-        print("Training voice cloning!")
+        print("Evaluating voice cloning!")
         fields += ["spk_emb"]
-        converters["spk_emb"] = np.load
     else:
         print("single speaker vits!")
     print("spk_num:", spk_num)

paddlespeech/t2s/exps/vits/voice_cloning.py

@@ -102,7 +102,7 @@ def voice_cloning(args):
         phone_ids = input_ids["phone_ids"][0]
     else:
         wav, _ = librosa.load(str(args.audio_path), sr=config.fs)
-        feats = spec_extractor.get_linear_spectrogram(wav)
+        feats = paddle.to_tensor(spec_extractor.get_linear_spectrogram(wav))
 
         mel_sequences = p.extract_mel_partials(
             p.preprocess_wav(args.audio_path))
@@ -122,10 +122,11 @@ def voice_cloning(args):
 
         with paddle.no_grad():
             if args.audio_path is None:
-                wav = vits.inference(text=phone_ids, spembs=spk_emb)
+                out = vits.inference(text=phone_ids, spembs=spk_emb)
             else:
-                wav = vits.voice_conversion(
+                out = vits.voice_conversion(
                     feats=feats, spembs_src=spk_emb_src, spembs_tgt=spk_emb)
+            wav = out["wav"]
 
         sf.write(
             str(output_dir / (utt_id + ".wav")),
@@ -138,10 +139,11 @@ def voice_cloning(args):
     utt_id = "random_spk_emb"
     with paddle.no_grad():
         if args.audio_path is None:
-            wav = vits.inference(text=phone_ids, spembs=random_spk_emb)
+            out = vits.inference(text=phone_ids, spembs=random_spk_emb)
         else:
-            wav = vits.voice_conversion(
+            out = vits.voice_conversion(
                 feats=feats, spembs_src=spk_emb_src, spembs_tgt=random_spk_emb)
+        wav = out["wav"]
     sf.write(
         str(output_dir / (utt_id + ".wav")), wav.numpy(), samplerate=config.fs)
     print(f"{utt_id} done!")

paddlespeech/t2s/models/vits/generator.py

@@ -524,8 +524,8 @@ class VITSGenerator(nn.Layer):
 
     def voice_conversion(
             self,
-            feats: Optional[paddle.Tensor]=None,
-            feats_lengths: Optional[paddle.Tensor]=None,
+            feats: paddle.Tensor=None,
+            feats_lengths: paddle.Tensor=None,
             sids_src: Optional[paddle.Tensor]=None,
             sids_tgt: Optional[paddle.Tensor]=None,
             spembs_src: Optional[paddle.Tensor]=None,

paddlespeech/t2s/models/vits/vits.py

@@ -381,7 +381,7 @@ class VITS(nn.Layer):
         if use_teacher_forcing:
             assert feats is not None
             feats = feats[None].transpose([0, 2, 1])
-            feats_lengths = paddle.to_tensor([paddle.shape(feats)[2]])
+            feats_lengths = paddle.to_tensor(paddle.shape(feats)[2])
             wav, att_w, dur = self.generator.inference(
                 text=text,
                 text_lengths=text_lengths,
@@ -409,7 +409,7 @@ class VITS(nn.Layer):
 
     def voice_conversion(
             self,
-            feats: Optional[paddle.Tensor]=None,
+            feats: paddle.Tensor,
             sids_src: Optional[paddle.Tensor]=None,
             sids_tgt: Optional[paddle.Tensor]=None,
             spembs_src: Optional[paddle.Tensor]=None,
@@ -429,7 +429,7 @@ class VITS(nn.Layer):
         """
         assert feats is not None
         feats = feats[None].transpose([0, 2, 1])
-        feats_lengths = paddle.to_tensor([paddle.shape(feats)[2]])
+        feats_lengths = paddle.to_tensor(paddle.shape(feats)[2])
 
         sids_none = sids_src is None and sids_tgt is None
         spembs_none = spembs_src is None and spembs_tgt is None