Canary <0.5s inference fix via padding

Signed-off-by: Piotr Żelasko <[email protected]>

nemo/collections/asr/models/aed_multitask_models.py

@@ -972,6 +972,8 @@ def _setup_transcribe_dataloader(self, config: Dict) -> 'torch.utils.data.DataLo
             'text_field': config.get('text_field', 'answer'),
             'lang_field': config.get('lang_field', 'target_lang'),
             'channel_selector': config.get('channel_selector', None),
+            'pad_min_duration': config.get('pad_min_duration', 1.0),
+            'pad_direction': config.get('pad_direction', 'both'),
         }
 
         temporary_datalayer = self._setup_dataloader_from_config(config=DictConfig(dl_config))

nemo/collections/common/data/lhotse/dataloader.py

@@ -135,6 +135,9 @@ class LhotseDataLoadingConfig:
     rir_enabled: bool = False
     rir_path: str | None = None  # str, must point to a lhotse RecordingSet manifest
     rir_prob: float = 0.5
+    #   f. Padding to a minimum duration. Examples shorter than this will be padded, others are unaffected.
+    pad_min_duration: Optional[float] = None
+    pad_direction: str = "right"  # "right" | "left" | "both" | "random"
 
     # 5. Other Lhotse options.
     text_field: str = "text"  # key to read the transcript from
@@ -257,6 +260,9 @@ def get_lhotse_dataloader_from_config(
             keep_excessive_supervisions=config.keep_excessive_supervisions,
         )
 
+    if config.pad_min_duration is not None:
+        cuts = cuts.pad(duration=config.pad_min_duration, direction=config.pad_direction, preserve_id=True)
+
     # Duration filtering, same as native NeMo dataloaders.
     # We can filter after the augmentations because they are applied only when calling load_audio().
     cuts = cuts.filter(DurationFilter(config.min_duration, config.max_duration))

tests/collections/common/test_lhotse_dataloading.py

@@ -23,7 +23,7 @@
 import torch
 from lhotse import CutSet, MonoCut, NumpyFilesWriter, Recording, SupervisionSegment, compute_num_samples
 from lhotse.audio import AudioLoadingError
-from lhotse.cut import Cut, MixedCut
+from lhotse.cut import Cut, MixedCut, PaddingCut
 from lhotse.cut.text import TextPairExample
 from lhotse.testing.dummies import dummy_recording
 from omegaconf import OmegaConf
@@ -311,6 +311,32 @@ def test_dataloader_from_lhotse_cuts_cut_into_windows(cutset_path: Path):
     # exactly 20 cuts were used because we cut 10x 1s cuts into 20x 0.5s cuts
 
 
+def test_dataloader_from_lhotse_cuts_pad_min_duration(cutset_path: Path):
+    config = OmegaConf.create(
+        {
+            "cuts_path": cutset_path,
+            "pad_min_duration": 21.0,
+            "pad_direction": "left",
+            "sample_rate": 16000,
+            "shuffle": True,
+            "use_lhotse": True,
+            "num_workers": 0,
+            "batch_size": 1,
+            "seed": 0,
+        }
+    )
+
+    dl = get_lhotse_dataloader_from_config(config=config, global_rank=0, world_size=1, dataset=Identity())
+
+    batch = next(iter(dl))
+    (cut,) = batch
+    assert cut.duration == 21.0
+    assert isinstance(cut, MixedCut)
+    assert len(cut.tracks) == 2
+    assert isinstance(cut.tracks[0].cut, PaddingCut)
+    assert isinstance(cut.tracks[1].cut, MonoCut)
+
+
 def test_dataloader_from_lhotse_cuts_channel_selector(mc_cutset_path: Path):
     # Dataloader without channel selector
     config = OmegaConf.create(