Adding pre-trained Overflow model (#2211)

* Adding pretrained Overflow model

* Stabilize HMM

* Fixup model manager

* Return `audio_unique_name` by default

* Distribute max split size over datasets

* Fixup eval_split_size

* Make style

TTS/.models.json

@@ -150,6 +150,15 @@
                     "author": "Eren Gölge @erogol",
                     "license": "apache 2.0",
                     "contact": "egolge@coqui.com"
+                },
+                "overflow": {
+                    "description": "Overflow model trained on LJSpeech",
+                    "github_rls_url": "https://coqui.gateway.scarf.sh/v0.10.0_models/tts_models--en--ljspeech--overflow.zip",
+                    "default_vocoder": "vocoder_models/en/ljspeech/hifigan_v2",
+                    "commit": "3b1a28f",
+                    "author": "Eren Gölge @erogol",
+                    "license": "apache 2.0",
+                    "contact": "egolge@coqui.ai"
                 }
             },
             "vctk": {

TTS/tts/datasets/__init__.py

@@ -129,7 +129,8 @@ def load_tts_samples(
                 meta_data_eval = formatter(root_path, meta_file_val, ignored_speakers=ignored_speakers)
                 meta_data_eval = add_extra_keys(meta_data_eval, language, dataset_name)
             else:
-                meta_data_eval, meta_data_train = split_dataset(meta_data_train, eval_split_max_size, eval_split_size)
+                eval_size_per_dataset = eval_split_max_size // len(datasets) if eval_split_max_size else None
+                meta_data_eval, meta_data_train = split_dataset(meta_data_train, eval_size_per_dataset, eval_split_size)
             meta_data_eval_all += meta_data_eval
         meta_data_train_all += meta_data_train
         # load attention masks for the duration predictor training

TTS/tts/datasets/dataset.py

@@ -520,6 +520,7 @@ class TTSDataset(Dataset):
                 "raw_text": batch["raw_text"],
                 "pitch": pitch,
                 "language_ids": language_ids,
+                "audio_unique_names": batch["audio_unique_name"],
             }
 
         raise TypeError(

TTS/tts/layers/overflow/neural_hmm.py

@@ -311,7 +311,7 @@ class NeuralHMM(nn.Module):
 
         # If the length of the mel is less than the number of states it will select the -inf values leading to nan gradients
         # Ideally, we should clean the dataset otherwise this is a little hack uncomment the line below
-        # final_log_c = final_log_c.clamp(min=torch.finfo(final_log_c.dtype).min)
+        final_log_c = final_log_c.clamp(min=torch.finfo(final_log_c.dtype).min)
 
         sum_final_log_c = torch.logsumexp(final_log_c, dim=1)
         return sum_final_log_c

TTS/tts/models/base_tts.py

@@ -232,6 +232,7 @@ class BaseTTS(BaseTrainerModel):
             "waveform": waveform,
             "pitch": pitch,
             "language_ids": language_ids,
+            "audio_unique_names": batch["audio_unique_names"],
         }
 
     def get_sampler(self, config: Coqpit, dataset: TTSDataset, num_gpus=1):
@@ -388,6 +389,9 @@ class BaseTTS(BaseTrainerModel):
         test_sentences = self.config.test_sentences
         aux_inputs = self._get_test_aux_input()
         for idx, sen in enumerate(test_sentences):
+            if isinstance(sen, list):
+                aux_inputs = self.get_aux_input_from_test_sentences(sen)
+                sen = aux_inputs["text"]
             outputs_dict = synthesis(
                 self,
                 sen,

TTS/utils/manage.py

@@ -366,6 +366,7 @@ class ModelManager(object):
         for file_path in z.namelist()[1:]:
             src_path = os.path.join(output_folder, file_path)
             dst_path = os.path.join(output_folder, os.path.basename(file_path))
+            if src_path != dst_path:
                 copyfile(src_path, dst_path)
         # remove the extracted folder
         rmtree(os.path.join(output_folder, z.namelist()[0]))