check for utterance len (#795)

Co-authored-by: behnam <basefisaray@roku.com>

egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py

@@ -1086,7 +1086,33 @@ def run(rank, world_size, args):
         # You should use ../local/display_manifest_statistics.py to get
         # an utterance duration distribution for your dataset to select
         # the threshold
-        return 1.0 <= c.duration <= 20.0
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
 
     train_cuts = train_cuts.filter(remove_short_and_long_utt)
 

egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py

@@ -1077,7 +1077,33 @@ def run(rank, world_size, args):
         # You should use ../local/display_manifest_statistics.py to get
         # an utterance duration distribution for your dataset to select
         # the threshold
-        return 1.0 <= c.duration <= 20.0
+        if c.duration < 1.0 or c.duration > 20.0:
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. Duration: {c.duration}"
+            )
+            return False
+
+        # In pruned RNN-T, we require that T >= S
+        # where T is the number of feature frames after subsampling
+        # and S is the number of tokens in the utterance
+
+        # In ./zipformer.py, the conv module uses the following expression
+        # for subsampling
+        T = ((c.num_frames - 7) // 2 + 1) // 2
+        tokens = sp.encode(c.supervisions[0].text, out_type=str)
+
+        if T < len(tokens):
+            logging.warning(
+                f"Exclude cut with ID {c.id} from training. "
+                f"Number of frames (before subsampling): {c.num_frames}. "
+                f"Number of frames (after subsampling): {T}. "
+                f"Text: {c.supervisions[0].text}. "
+                f"Tokens: {tokens}. "
+                f"Number of tokens: {len(tokens)}"
+            )
+            return False
+
+        return True
 
     train_cuts = train_cuts.filter(remove_short_and_long_utt)