Fixed bug related to yourtts speaker embeddings issue (#2234)

* Fixed bug related to yourtts speaker embeddings issue

* Reverted code for base_tts

* Bug fix on VITS d_vector_file type

* Ignore the test speakers on YourTTS recipe

* Add speaker encoder model and config on YourTTS recipe to easily do zero-shot inference

* Update YourTTS config file

* Update ModelManager._update_path to deal with list attributes

* Fix lint checks

* Remove unused code

* Fix unit tests

* Reset name_to_id to get the right speaker ids on load_embeddings_from_list_of_files

* Set weighted_sampler_multipliers as an empty dict to prevent users' mistakes

Co-authored-by: Edresson Casanova <edresson1@gmail.com>

TTS/.models.json

@@ -4,7 +4,7 @@
             "multi-dataset":{
                 "your_tts":{
                     "description": "Your TTS model accompanying the paper https://arxiv.org/abs/2112.02418",
-                    "github_rls_url": "https://coqui.gateway.scarf.sh/v0.6.1_models/tts_models--multilingual--multi-dataset--your_tts.zip",
+                    "github_rls_url": "https://coqui.gateway.scarf.sh/v0.10.1_models/tts_models--multilingual--multi-dataset--your_tts.zip",
                     "default_vocoder": null,
                     "commit": "e9a1953e",
                     "license": "CC BY-NC-ND 4.0",

TTS/tts/configs/vits_config.py

@@ -167,7 +167,7 @@ class VitsConfig(BaseTTSConfig):
 
     # use d-vectors
     use_d_vector_file: bool = False
-    d_vector_file: str = None
+    d_vector_file: List[str] = None
     d_vector_dim: int = None
 
     def __post_init__(self):

TTS/tts/models/vits.py

@@ -477,8 +477,8 @@ class VitsArgs(Coqpit):
         use_d_vector_file (bool):
             Enable/Disable the use of d-vectors for multi-speaker training. Defaults to False.
 
-        d_vector_file (str):
-            Path to the file including pre-computed speaker embeddings. Defaults to None.
+        d_vector_file (List[str]):
+            List of paths to the files including pre-computed speaker embeddings. Defaults to None.
 
         d_vector_dim (int):
             Number of d-vector channels. Defaults to 0.
@@ -573,7 +573,7 @@ class VitsArgs(Coqpit):
     use_speaker_embedding: bool = False
     num_speakers: int = 0
     speakers_file: str = None
-    d_vector_file: str = None
+    d_vector_file: List[str] = None
     speaker_embedding_channels: int = 256
     use_d_vector_file: bool = False
     d_vector_dim: int = 0

TTS/tts/utils/managers.py

@@ -235,6 +235,9 @@ class EmbeddingManager(BaseIDManager):
             self.embeddings_by_names.update(embeddings_by_names)
             self.embeddings.update(embeddings)
 
+        # reset name_to_id to get the right speaker ids
+        self.name_to_id = {name: i for i, name in enumerate(self.name_to_id)}
+
     def get_embedding_by_clip(self, clip_idx: str) -> List:
         """Get embedding by clip ID.
 

TTS/tts/utils/speakers.py

@@ -109,10 +109,6 @@ class SpeakerManager(EmbeddingManager):
 
         if get_from_config_or_model_args_with_default(config, "use_d_vector_file", False):
             speaker_manager = SpeakerManager()
-            if get_from_config_or_model_args_with_default(config, "speakers_file", None):
-                speaker_manager = SpeakerManager(
-                    d_vectors_file_path=get_from_config_or_model_args_with_default(config, "speaker_file", None)
-                )
             if get_from_config_or_model_args_with_default(config, "d_vector_file", None):
                 speaker_manager = SpeakerManager(
                     d_vectors_file_path=get_from_config_or_model_args_with_default(config, "d_vector_file", None)

TTS/utils/manage.py

@@ -339,10 +339,18 @@ class ModelManager(object):
                         sub_conf = sub_conf[fd]
                     else:
                         return
-                sub_conf[field_names[-1]] = new_path
+                if isinstance(sub_conf[field_names[-1]], list):
+                    sub_conf[field_names[-1]] = [new_path]
+                else:
+                    sub_conf[field_names[-1]] = new_path
             else:
                 # field name points to a top-level field
-                config[field_name] = new_path
+                if not field_name in config:
+                    return
+                if isinstance(config[field_name], list):
+                    config[field_name] = [new_path]
+                else:
+                    config[field_name] = new_path
             config.save_json(config_path)
 
     @staticmethod

recipes/vctk/yourtts/train_yourtts.py

@@ -57,7 +57,25 @@ if not os.path.exists(VCTK_DOWNLOAD_PATH):
 
 # init configs
 vctk_config = BaseDatasetConfig(
-    formatter="vctk", dataset_name="vctk", meta_file_train="", meta_file_val="", path=VCTK_DOWNLOAD_PATH, language="en"
+    formatter="vctk",
+    dataset_name="vctk",
+    meta_file_train="",
+    meta_file_val="",
+    path=VCTK_DOWNLOAD_PATH,
+    language="en",
+    ignored_speakers=[
+        "p261",
+        "p225",
+        "p294",
+        "p347",
+        "p238",
+        "p234",
+        "p248",
+        "p335",
+        "p245",
+        "p326",
+        "p302",
+    ],  # Ignore the test speakers to full replicate the paper experiment
 )
 
 # Add here all datasets configs, in our case we just want to train with the VCTK dataset then we need to add just VCTK. Note: If you want to added new datasets just added they here and it will automatically compute the speaker embeddings (d-vectors) for this new dataset :)
@@ -111,11 +129,11 @@ model_args = VitsArgs(
     use_d_vector_file=True,
     d_vector_dim=512,
     num_layers_text_encoder=10,
+    speaker_encoder_model_path=SPEAKER_ENCODER_CHECKPOINT_PATH,
+    speaker_encoder_config_path=SPEAKER_ENCODER_CONFIG_PATH,
     resblock_type_decoder="2",  # On the paper, we accidentally trained the YourTTS using ResNet blocks type 2, if you like you can use the ResNet blocks type 1 like the VITS model
     # Usefull parameters to enable the Speaker Consistency Loss (SCL) discribed in the paper
     # use_speaker_encoder_as_loss=True,
-    # speaker_encoder_model_path=SPEAKER_ENCODER_CHECKPOINT_PATH,
-    # speaker_encoder_config_path=SPEAKER_ENCODER_CONFIG_PATH,
     # Usefull parameters to the enable multilingual training
     # use_language_embedding=True,
     # embedded_language_dim=4,
@@ -207,6 +225,7 @@ config = VitsConfig(
     use_weighted_sampler=True,
     # Ensures that all speakers are seen in the training batch equally no matter how many samples each speaker has
     weighted_sampler_attrs={"speaker_name": 1.0},
+    weighted_sampler_multipliers={},
     # It defines the Speaker Consistency Loss (SCL) α to 9 like the paper
     speaker_encoder_loss_alpha=9.0,
 )

tests/tts_tests/test_vits.py

@@ -210,7 +210,7 @@ class TestVits(unittest.TestCase):
             num_chars=32,
             use_d_vector_file=True,
             d_vector_dim=256,
-            d_vector_file=os.path.join(get_tests_data_path(), "dummy_speakers.json"),
+            d_vector_file=[os.path.join(get_tests_data_path(), "dummy_speakers.json")],
         )
         config = VitsConfig(model_args=args)
         model = Vits.init_from_config(config, verbose=False).to(device)
@@ -355,7 +355,7 @@ class TestVits(unittest.TestCase):
             num_chars=32,
             use_d_vector_file=True,
             d_vector_dim=256,
-            d_vector_file=os.path.join(get_tests_data_path(), "dummy_speakers.json"),
+            d_vector_file=[os.path.join(get_tests_data_path(), "dummy_speakers.json")],
         )
         config = VitsConfig(model_args=args)
         model = Vits.init_from_config(config, verbose=False).to(device)
@@ -587,7 +587,7 @@ class TestVits(unittest.TestCase):
                 num_chars=32,
                 use_d_vector_file=True,
                 d_vector_dim=256,
-                d_vector_file=os.path.join(get_tests_data_path(), "dummy_speakers.json"),
+                d_vector_file=[os.path.join(get_tests_data_path(), "dummy_speakers.json")],
             )
         )
         model = Vits.init_from_config(config, verbose=False).to(device)

tests/tts_tests/test_vits_d-vectors_train.py

@@ -33,7 +33,7 @@ config.audio.trim_db = 60
 
 # active multispeaker d-vec mode
 config.model_args.use_d_vector_file = True
-config.model_args.d_vector_file = "tests/data/ljspeech/speakers.json"
+config.model_args.d_vector_file = ["tests/data/ljspeech/speakers.json"]
 config.model_args.d_vector_dim = 256
 
 

tests/tts_tests/test_vits_multilingual_train-d_vectors.py

@@ -63,8 +63,8 @@ config.use_speaker_embedding = False
 # active multispeaker d-vec mode
 config.model_args.use_d_vector_file = True
 config.use_d_vector_file = True
-config.model_args.d_vector_file = "tests/data/ljspeech/speakers.json"
-config.d_vector_file = "tests/data/ljspeech/speakers.json"
+config.model_args.d_vector_file = ["tests/data/ljspeech/speakers.json"]
+config.d_vector_file = ["tests/data/ljspeech/speakers.json"]
 config.model_args.d_vector_dim = 256
 config.d_vector_dim = 256