Updated model, created python package, and released LJ-Speech model

README.md

@@ -56,11 +56,11 @@ tar -xvjf LJSpeech-1.1.tar.bz2
 4. Download the train split here and extract it in the root directory of the repo. 
 5. Extract Mel spectrograms and preprocess audio:
 ```
-python preprocess.py in_dir=path/to/LJSpeech-1.1
+python preprocess.py in_dir=path/to/LJSpeech-1.1 out_dir=datasets/LJSpeech-1.1
 ```
 6. Train the model:
 ```
-python train.py checkpoint_dir=ljspeech
+python train.py checkpoint_dir=ljspeech dataset_dir=datasets/LJSpeech-1.1
 ```
 
 ## Pretrained Models

preprocess.py

@@ -81,7 +81,7 @@ def process_wav(wav_path, out_path, cfg):
 @hydra.main(config_path="univoc/config", config_name="preprocess")
 def preprocess_dataset(cfg):
     in_dir = Path(utils.to_absolute_path(cfg.in_dir))
-    out_dir = Path(utils.to_absolute_path("datasets"))
+    out_dir = Path(utils.to_absolute_path(cfg.out_dir))
     out_dir.mkdir(parents=True, exist_ok=True)
 
     executor = ProcessPoolExecutor(max_workers=cpu_count())

requirements.txt

@@ -2,3 +2,5 @@ numpy>=1.19.2
 librosa>=0.8.0
 tqdm>=4.50
 hydra-core>=1.0.3
+pyloudnorm>=0.1.0
+tensorboard>=2.3.0

train.py

@@ -67,7 +67,7 @@ def train_model(cfg):
     else:
         global_step = 0
 
-    dataset_root = Path(utils.to_absolute_path("datasets"))
+    dataset_root = Path(utils.to_absolute_path(cfg.dataset_dir))
     dataset = VocoderDataset(
         dataset_root,
         sample_frames=cfg.train.sample_frames,

univoc/config/preprocess.yaml

@@ -1,4 +1,5 @@
 defaults:
   - config
 
-in_dir: ???
+in_dir: ???
+out_dir: ???

univoc/config/train.yaml

@@ -2,4 +2,5 @@ defaults:
   - config
 
 resume: false
-checkpoint_dir: ???
+checkpoint_dir: ???
+dataset_dir: ???