Reazon speech - add streaming

egs/librispeech/ASR/zipformer/decode.py

@@ -122,6 +122,7 @@
     modified_beam_search_LODR,
 )
 from lhotse import set_caching_enabled
+from tokenizer import Tokenizer
 from train import add_model_arguments, get_model, get_params
 
 from icefall import ContextGraph, LmScorer, NgramLm
@@ -377,6 +378,17 @@ def get_parser():
         default=False,
         help="""Skip scoring, but still save the ASR output (for eval sets).""",
     )
+    parser.add_argument(
+        "--blank-penalty",
+        type=float,
+        default=0.0,
+        help="""
+        The penalty applied on blank symbol during decoding.
+        Note: It is a positive value that would be applied to logits like
+        this `logits[:, 0] -= blank_penalty` (suppose logits.shape is
+        [batch_size, vocab] and blank id is 0).
+        """,
+    )
 
     add_model_arguments(parser)
 
@@ -601,6 +613,7 @@ def decode_one_batch(
 
     # prefix = ( "greedy_search" | "fast_beam_search_nbest" | "modified_beam_search" )
     prefix = f"{params.decoding_method}"
+    key = f"blank_penalty_{params.blank_penalty}"
     if params.decoding_method == "greedy_search":
         return {"greedy_search": hyps}
     elif "fast_beam_search" in params.decoding_method:

egs/librispeech/ASR/zipformer/zipformer.py

@@ -2434,4 +2434,4 @@ def _test_zipformer_main(causal: bool = False):
     torch.set_num_threads(1)
     torch.set_num_interop_threads(1)
     _test_zipformer_main(False)
-    _test_zipformer_main(True)
+    _test_zipformer_main(True)

egs/reazonspeech/ASR/RESULTS.md

@@ -47,3 +47,41 @@ The decoding command is:
     --blank-penalty 0
 ```
 
+#### Streaming
+
+We have not completed evaluation of our models yet and will add evaluation results here once it's completed.
+
+The training command is:
+```shell
+./zipformer/train.py \
+  --world-size 8 \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir zipformer/exp-large \
+  --causal 1 \
+  --num-encoder-layers 2,2,4,5,4,2 \
+  --feedforward-dim 512,768,1536,2048,1536,768 \
+  --encoder-dim 192,256,512,768,512,256 \
+  --encoder-unmasked-dim 192,192,256,320,256,192 \
+  --lang data/lang_char \
+  --max-duration 1600
+```
+
+The decoding command is:
+
+```shell
+./zipformer/streaming_decode.py \
+  --epoch 28 \
+  --avg 15 \
+  --causal 1 \
+  --chunk-size 32 \
+  --left-context-frames 256 \
+  --exp-dir ./zipformer/exp-large \
+  --lang data/lang_char \
+  --num-encoder-layers 2,2,4,5,4,2 \
+  --feedforward-dim 512,768,1536,2048,1536,768 \
+  --encoder-dim 192,256,512,768,512,256 \
+  --encoder-unmasked-dim 192,192,256,320,256,192
+```
+

egs/reazonspeech/ASR/local/utils/tokenizer.py

@@ -12,7 +12,6 @@ class Tokenizer:
     @staticmethod
     def add_arguments(parser: argparse.ArgumentParser):
         group = parser.add_argument_group(title="Lang related options")
-
         group.add_argument("--lang", type=Path, help="Path to lang directory.")
 
         group.add_argument(