Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

fix(kws): ignore empty keyword entries in KwsCtcPrefixDecoder (#3800)

Mohammad Hijjawic9c42c0bb7dd
diff --git a/funasr/utils/kws_utils.py b/funasr/utils/kws_utils.py
index 11de3238b..59447361f 100644
--- a/funasr/utils/kws_utils.py
+++ b/funasr/utils/kws_utils.py
@@ -117,6 +117,11 @@ class KwsCtcPrefixDecoder():
keywords_list = self.keywords_str.strip().replace(' ', '').split(',')
for keyword in keywords_list:
strs, indexs = query_token_set(keyword, token_table, seg_dict)
+ if len(indexs) == 0:
+ # An empty entry (e.g. from a trailing comma) would match every
+ # hypothesis, so skip it instead of reporting a false detection.
+ logging.warning(f"Skipping keyword {keyword!r}: it has no tokens to match.")
+ continue
self.keywords_token[keyword] = {}
self.keywords_token[keyword]['token_id'] = indexs
self.keywords_token[keyword]['token_str'] = ''.join('%s ' % str(i) for i in indexs)
diff --git a/tests/test_kws_utils.py b/tests/test_kws_utils.py
index 79fe4cdc8..12e8fd1a8 100644
--- a/tests/test_kws_utils.py
+++ b/tests/test_kws_utils.py
@@ -1,7 +1,27 @@
import re
import unittest
-from funasr.utils.kws_utils import query_token_set, split_mixed_label, symbol_str
+import torch
+
+from funasr.utils.kws_utils import (
+ KwsCtcPrefixDecoder,
+ query_token_set,
+ split_mixed_label,
+ symbol_str,
+)
+
+
+class _IdentityCtc:
+ def softmax(self, x):
+ return x
+
+
+def _posteriors(token_ids, vocab_size=4):
+ # One frame per entry: token_id gets all the probability mass (0 is blank).
+ probs = torch.zeros(len(token_ids), vocab_size)
+ for frame, token_id in enumerate(token_ids):
+ probs[frame, token_id] = 1.0
+ return probs
class KwsUtilsTest(unittest.TestCase):
@@ -18,6 +38,19 @@ class KwsUtilsTest(unittest.TestCase):
self.assertEqual(tokens, ("a", "中"))
self.assertEqual(token_ids, (1, 2))
+ def test_empty_keyword_entries_do_not_produce_detections(self):
+ token_list = ["<blank>", "<unk>", "小", "云"]
+ for keywords in ("小云小云,", "小云小云, ,", ",小云小云", "小云小云,!!"):
+ with self.subTest(keywords=keywords):
+ decoder = KwsCtcPrefixDecoder(_IdentityCtc(), keywords, token_list, {})
+
+ self.assertEqual(decoder.decode(_posteriors([0, 0, 0, 0])), (False, None, None))
+ self.assertEqual(decoder.decode(_posteriors([2, 0, 0, 0])), (False, None, None))
+ self.assertEqual(
+ decoder.decode(_posteriors([2, 3, 0, 2, 3])), (True, "小云小云", 1.0)
+ )
+ self.assertEqual(list(decoder.keywords_token), ["小云小云"])
+
if __name__ == "__main__":
unittest.main()