update

egs/gigaspeech/ASR/local/compute_fbank_gigaspeech_splits.py

@@ -22,7 +22,12 @@
 from pathlib import Path
 
 import torch
-from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig
+from lhotse import (
+    CutSet,
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    set_audio_duration_mismatch_tolerance,
+)
 
 # Torch's multithreaded behavior needs to be disabled or
 # it wastes a lot of CPU and slow things down.
@@ -80,7 +85,7 @@ def compute_fbank_gigaspeech_splits(args):
     output_dir = Path(output_dir)
     assert output_dir.exists(), f"{output_dir} does not exist!"
 
-    num_digits = 8  # num_digits is fixed by lhotse split-lazy
+    num_digits = 4
 
     start = args.start
     stop = args.stop
@@ -95,8 +100,10 @@ def compute_fbank_gigaspeech_splits(args):
     extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
     logging.info(f"device: {device}")
 
+    set_audio_duration_mismatch_tolerance(0.1)
+
     for i in range(start, stop):
-        idx = f"{i + 1}".zfill(num_digits)
+        idx = f"{i}".zfill(num_digits)
         logging.info(f"Processing {idx}/{num_splits}")
 
         cuts_path = output_dir / f"gigaspeech_cuts_XL.{idx}.jsonl.gz"

egs/gigaspeech/ASR/local/preprocess_gigaspeech.py

@@ -71,7 +71,7 @@ def preprocess_giga_speech():
 
     for partition, m in manifests.items():
         logging.info(f"Processing {partition}")
-        raw_cuts_path = output_dir / f"cuts_{partition}_raw.jsonl.gz"
+        raw_cuts_path = output_dir / f"gigaspeech_cuts_{partition}_raw.jsonl.gz"
         if raw_cuts_path.is_file():
             logging.info(f"{partition} already exists - skipping")
             continue
@@ -91,14 +91,6 @@ def preprocess_giga_speech():
             recordings=m["recordings"],
             supervisions=m["supervisions"],
         )
-        # Run data augmentation that needs to be done in the
-        # time domain.
-        if partition not in ["DEV", "TEST"]:
-            logging.info(
-                f"Speed perturb for {partition} with factors 0.9 and 1.1 "
-                "(Perturbing may take 8 minutes and saving may take 20 minutes)"
-            )
-            cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
         logging.info(f"Saving to {raw_cuts_path}")
         cut_set.to_file(raw_cuts_path)
 

egs/gigaspeech/ASR/prepare.sh

@@ -9,9 +9,9 @@ nj=16
 stage=0
 stop_stage=100
 
-# Split XL subset to a number of pieces (about 2000)
+# Split XL subset to a number of pieces
 # This is to avoid OOM during feature extraction.
-num_per_split=50
+num_splits=1000
 
 # We assume dl_dir (download dir) contains the following
 # directories and files. If not, they will be downloaded
@@ -122,7 +122,7 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
     $dl_dir/GigaSpeech data/manifests
 fi
 
-if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+if 0 && [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
   log "Stage 2: Prepare musan manifest"
   # We assume that you have downloaded the musan corpus
   # to $dl_dir/musan
@@ -147,7 +147,7 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
   log "Stage 5: Split XL subset into pieces (may take 30 minutes)"
   split_dir=data/fbank/XL_split
   if [ ! -f $split_dir/.split_completed ]; then
-    lhotse split-lazy ./data/fbank/gigapspeech_cuts_XL_raw.jsonl.gz $split_dir $num_per_split
+    lhotse split -s $num_splits ./data/fbank/gigaspeech_cuts_XL_raw.jsonl.gz $split_dir
     touch $split_dir/.split_completed
   fi
 fi
@@ -161,7 +161,7 @@ if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
     --num-splits $num_splits
 fi
 
-if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+if 0 && [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
   log "Stage 7: Compute fbank for musan"
   mkdir -p data/fbank
   ./local/compute_fbank_musan.py

egs/gigaspeech/ASR/zipformer/asr_datamodule.py

@@ -194,6 +194,13 @@ def add_arguments(cls, parser: argparse.ArgumentParser):
             "with training dataset. ",
         )
 
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
         # GigaSpeech specific arguments
         group.add_argument(
             "--subset",
@@ -209,9 +216,7 @@ def add_arguments(cls, parser: argparse.ArgumentParser):
         )
 
     def train_dataloaders(
-        self,
-        cuts_train: CutSet,
-        sampler_state_dict: Optional[Dict[str, Any]] = None,
+        self, cuts_train: CutSet, sampler_state_dict: Optional[Dict[str, Any]] = None,
     ) -> DataLoader:
         """
         Args:
@@ -353,13 +358,10 @@ def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
             )
         else:
             validate = K2SpeechRecognitionDataset(
-                cut_transforms=transforms,
-                return_cuts=self.args.return_cuts,
+                cut_transforms=transforms, return_cuts=self.args.return_cuts,
             )
         valid_sampler = DynamicBucketingSampler(
-            cuts_valid,
-            max_duration=self.args.max_duration,
-            shuffle=False,
+            cuts_valid, max_duration=self.args.max_duration, shuffle=False,
         )
         logging.info("About to create dev dataloader")
         valid_dl = DataLoader(
@@ -381,16 +383,11 @@ def test_dataloaders(self, cuts: CutSet) -> DataLoader:
             return_cuts=self.args.return_cuts,
         )
         sampler = DynamicBucketingSampler(
-            cuts,
-            max_duration=self.args.max_duration,
-            shuffle=False,
+            cuts, max_duration=self.args.max_duration, shuffle=False,
         )
         logging.debug("About to create test dataloader")
         test_dl = DataLoader(
-            test,
-            batch_size=None,
-            sampler=sampler,
-            num_workers=self.args.num_workers,
+            test, batch_size=None, sampler=sampler, num_workers=self.args.num_workers,
         )
         return test_dl