Fix VITS model SPD

TTS/tts/configs/vits_config.py

@@ -119,8 +119,8 @@ class VitsConfig(BaseTTSConfig):
     compute_linear_spec: bool = True
 
     # overrides
-    min_seq_len: int = 13
-    max_seq_len: int = 500
+    min_seq_len: int = 32
+    max_seq_len: int = 1000
     r: int = 1  # DO NOT CHANGE
     add_blank: bool = True
 

TTS/tts/layers/vits/stochastic_duration_predictor.py

@@ -228,7 +228,7 @@ class StochasticDurationPredictor(nn.Module):
             h = self.post_pre(dr)
             h = self.post_convs(h, x_mask)
             h = self.post_proj(h) * x_mask
-            noise = torch.rand(dr.size(0), 2, dr.size(2)).to(device=x.device, dtype=x.dtype) * x_mask
+            noise = torch.randn(dr.size(0), 2, dr.size(2)).to(device=x.device, dtype=x.dtype) * x_mask
             z_q = noise
 
             # posterior encoder

TTS/tts/models/base_tts.py

@@ -194,7 +194,7 @@ class BaseTTS(BaseModel):
             if hasattr(self, "make_symbols"):
                 custom_symbols = self.make_symbols(self.config)
 
-            # init dataloader
+            # init dataset
             dataset = TTSDataset(
                 outputs_per_step=config.r if "r" in config else 1,
                 text_cleaner=config.text_cleaner,
@@ -220,13 +220,15 @@ class BaseTTS(BaseModel):
                 else None,
             )
 
+            # pre-compute phonemes
             if config.use_phonemes and config.compute_input_seq_cache and rank in [None, 0]:
                 if hasattr(self, "eval_data_items") and is_eval:
                     dataset.items = self.eval_data_items
                 elif hasattr(self, "train_data_items") and not is_eval:
                     dataset.items = self.train_data_items
                 else:
-                    # precompute phonemes to have a better estimate of sequence lengths.
+                    # precompute phonemes for precise estimate of sequence lengths.
+                    # otherwise `dataset.sort_items()` uses raw text lengths
                     dataset.compute_input_seq(config.num_loader_workers)
 
                     # TODO: find a more efficient solution
@@ -240,9 +242,13 @@ class BaseTTS(BaseModel):
             if num_gpus > 1:
                 dist.barrier()
 
+            # sort input sequences from short to long
             dataset.sort_items()
 
+            # sampler for DDP
             sampler = DistributedSampler(dataset) if num_gpus > 1 else None
+
+            # init dataloader
             loader = DataLoader(
                 dataset,
                 batch_size=config.eval_batch_size if is_eval else config.batch_size,

TTS/tts/models/vits.py

@@ -177,7 +177,7 @@ class VitsArgs(Coqpit):
     num_layers_text_encoder: int = 6
     kernel_size_text_encoder: int = 3
     dropout_p_text_encoder: int = 0.1
-    dropout_p_duration_predictor: int = 0.1
+    dropout_p_duration_predictor: int = 0.5
     kernel_size_posterior_encoder: int = 5
     dilation_rate_posterior_encoder: int = 1
     num_layers_posterior_encoder: int = 16