add wavernn in synthesize_e2e, test=tts

examples/csmsc/tts3/local/synthesize_e2e.sh

@@ -89,3 +89,24 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
         --inference_dir=${train_output_path}/inference \
         --phones_dict=dump/phone_id_map.txt
 fi
+
+
+# wavernn
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    echo "in wavernn syn_e2e"
+    FLAGS_allocator_strategy=naive_best_fit \
+    FLAGS_fraction_of_gpu_memory_to_use=0.01 \
+    python3 ${BIN_DIR}/../synthesize_e2e.py \
+        --am=fastspeech2_csmsc \
+        --am_config=${config_path} \
+        --am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
+        --am_stat=dump/train/speech_stats.npy \
+        --voc=wavernn_csmsc \
+        --voc_config=wavernn_test/default.yaml \
+        --voc_ckpt=wavernn_test/snapshot_iter_5000.pdz \
+        --voc_stat=wavernn_test/feats_stats.npy \
+        --lang=zh \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/test_e2e \
+        --phones_dict=dump/phone_id_map.txt
+fi

paddlespeech/t2s/exps/synthesize_e2e.py

@@ -59,6 +59,10 @@ model_alias = {
     "paddlespeech.t2s.models.hifigan:HiFiGANGenerator",
     "hifigan_inference":
     "paddlespeech.t2s.models.hifigan:HiFiGANInference",
+    "wavernn":
+    "paddlespeech.t2s.models.wavernn:WaveRNN",
+    "wavernn_inference":
+    "paddlespeech.t2s.models.wavernn:WaveRNNInference",
 }
 
 
@@ -148,10 +152,16 @@ def evaluate(args):
     voc_name = args.voc[:args.voc.rindex('_')]
     voc_class = dynamic_import(voc_name, model_alias)
     voc_inference_class = dynamic_import(voc_name + '_inference', model_alias)
-    voc = voc_class(**voc_config["generator_params"])
-    voc.set_state_dict(paddle.load(args.voc_ckpt)["generator_params"])
-    voc.remove_weight_norm()
-    voc.eval()
+    if voc_name != 'wavernn':
+        voc = voc_class(**voc_config["generator_params"])
+        voc.set_state_dict(paddle.load(args.voc_ckpt)["generator_params"])
+        voc.remove_weight_norm()
+        voc.eval()
+    else:
+        voc = voc_class(**voc_config["model"])
+        voc.set_state_dict(paddle.load(args.voc_ckpt)["main_params"])
+        voc.eval()
+
     voc_mu, voc_std = np.load(args.voc_stat)
     voc_mu = paddle.to_tensor(voc_mu)
     voc_std = paddle.to_tensor(voc_std)
@@ -307,7 +317,8 @@ def main():
         default='pwgan_csmsc',
         choices=[
             'pwgan_csmsc', 'pwgan_ljspeech', 'pwgan_aishell3', 'pwgan_vctk',
-            'mb_melgan_csmsc', 'style_melgan_csmsc', 'hifigan_csmsc'
+            'mb_melgan_csmsc', 'style_melgan_csmsc', 'hifigan_csmsc',
+            'wavernn_csmsc'
         ],
         help='Choose vocoder type of tts task.')
 

paddlespeech/t2s/models/wavernn/wavernn.py

@@ -590,3 +590,27 @@ class WaveRNN(nn.Layer):
         for i in range(size):
             bar += '█' if i <= done else '░'
         return bar
+
+
+class WaveRNNInference(nn.Layer):
+    def __init__(self, normalizer, wavernn):
+        super().__init__()
+        self.normalizer = normalizer
+        self.wavernn = wavernn
+
+    def forward(self,
+                logmel,
+                batched: bool=True,
+                target: int=12000,
+                overlap: int=600,
+                mu_law: bool=True,
+                gen_display: bool=False):
+        normalized_mel = self.normalizer(logmel)
+        wav = self.wavernn.generate(
+            normalized_mel,
+            batched=batched,
+            target=target,
+            overlap=overlap,
+            mu_law=mu_law,
+            gen_display=gen_display)
+        return wav