New noise_norm perturbation based on Riva work

nemo/collections/asr/data/audio_to_text.py

@@ -191,8 +191,8 @@ def expand_sharded_filepaths(sharded_filepaths, shard_strategy: str, world_size:
                 sharded_filepaths = sharded_filepaths.replace(bkey, "}")
 
     if isinstance(sharded_filepaths, str):
-        # Brace expand
-        sharded_filepaths = list(braceexpand.braceexpand(sharded_filepaths))
+        # Brace expand, set escape=False for Windows compatibility
+        sharded_filepaths = list(braceexpand.braceexpand(sharded_filepaths, escape=False))
 
     # Expand store paths into WebDataset URLs
     sharded_filepaths = [
@@ -1359,5 +1359,9 @@ def __iter__(self):
         for dataset_idx in shuffled_order:
             d = self.datasets[dataset_idx]
             assert isinstance(d, IterableDataset), "ChainDataset only supports IterableDataset"
-            for x in d:
+            for idx, x in enumerate(d):
                 yield x
+                # in case d is an infinite dataset, we want to break the loop
+                # so that the other datasets get a chance to yield too
+                if idx >= len(d) - 1:
+                    break

nemo/collections/asr/data/audio_to_text_dataset.py

@@ -512,7 +512,7 @@ def get_audio_to_text_char_dataset_from_config(
         constructed dataset or None if dataset config is invalid or nothing to load
     """
     if 'augmentor' in config:
-        augmentor = process_augmentations(config['augmentor'])
+        augmentor = process_augmentations(config['augmentor'], global_rank=global_rank, world_size=world_size)
     else:
         augmentor = None
 
@@ -609,7 +609,7 @@ def get_audio_to_text_bpe_dataset_from_config(
         constructed dataset or None if dataset config is invalid or nothing to load
     """
     if 'augmentor' in config:
-        augmentor = process_augmentations(config['augmentor'])
+        augmentor = process_augmentations(config['augmentor'], global_rank=global_rank, world_size=world_size)
     else:
         augmentor = None
 

nemo/collections/asr/parts/preprocessing/__init__.py

@@ -20,6 +20,7 @@
     GainPerturbation,
     ImpulsePerturbation,
     NoisePerturbation,
+    NoisePerturbationWithNormalization,
     Perturbation,
     RirAndNoisePerturbation,
     ShiftPerturbation,