Skip to content

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

fix(hotwords): stop fuzzy postprocess matching from rewriting nearby text (#3744)

Chi-Wei Lee43329e8cefb4
diff --git a/funasr/utils/postprocess_hotwords.py b/funasr/utils/postprocess_hotwords.py
index 26e482697..5ff6183fa 100644
--- a/funasr/utils/postprocess_hotwords.py
+++ b/funasr/utils/postprocess_hotwords.py
@@ -19,6 +19,9 @@ HotwordInput = Union[str, Sequence[str], Mapping[str, str], None]
_EXPLICIT_SEPARATORS = ("=>", "->", "→")
_TOKEN_PATTERN = re.compile(r"[\u4e00-\u9fff]|[a-zA-Z]+|[0-9]+")
+_HAN_PATTERN = re.compile(r"[\u4e00-\u9fff]")
+_LATIN_PATTERN = re.compile(r"[a-zA-Z]")
+_ASCII_ALNUM = frozenset("abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789")
_LAZY_PINYIN = None
_PINYIN_STYLE = None
@@ -76,7 +79,10 @@ def _require_rapidfuzz():
def _to_pinyin_key(text: str) -> str:
lazy_pinyin, style = _require_pypinyin()
- return "".join(lazy_pinyin(text, style=style.NORMAL, errors="ignore")).lower()
+ # Keep non-Chinese characters in the key. Dropping them makes a window with
+ # extra punctuation, digits or Latin letters score 1.0 against the target,
+ # and gives every Latin-only target and window the same empty key.
+ return "".join(lazy_pinyin(text, style=style.NORMAL, errors="default")).lower()
def _parse_line(line: str) -> Tuple[Optional[str], Optional[str], bool]:
@@ -230,14 +236,21 @@ class PostprocessHotwordMatcher:
seen.add(target_s)
self.fuzzy_targets.append(target_s)
- self._length_buckets: Dict[int, List[Tuple[str, str]]] = {}
+ self._length_buckets: Dict[int, List[Tuple[str, str, bool, bool]]] = {}
self._fuzz = None
if self.fuzzy_targets and self.enable_fuzzy:
self._fuzz = _require_rapidfuzz()
_require_pypinyin()
for target in self.fuzzy_targets:
bucket = self._length_buckets.setdefault(len(target), [])
- bucket.append((target, _to_pinyin_key(target)))
+ bucket.append(
+ (
+ target,
+ _to_pinyin_key(target),
+ bool(_HAN_PATTERN.search(target)),
+ bool(_LATIN_PATTERN.search(target)),
+ )
+ )
def apply_text(self, text: str) -> Tuple[str, List[HotwordMatch]]:
if not text:
@@ -330,12 +343,27 @@ class PostprocessHotwordMatcher:
segment = text[start:end]
if not segment or not _TOKEN_PATTERN.search(segment):
continue
+ # Do not split a Latin word or a number: with target "Win11",
+ # the window "Win1" inside "Win10" would become "Win110".
+ if start > 0 and text[start - 1] in _ASCII_ALNUM and text[start] in _ASCII_ALNUM:
+ continue
+ if end < text_len and text[end - 1] in _ASCII_ALNUM and text[end] in _ASCII_ALNUM:
+ continue
segment_py = _to_pinyin_key(segment)
+ segment_han = bool(_HAN_PATTERN.search(segment))
+ segment_latin = bool(_LATIN_PATTERN.search(segment))
for length in bucket_keys:
- for target, target_py in self._length_buckets[length]:
+ for target, target_py, target_han, target_latin in self._length_buckets[length]:
if segment == target:
continue
+ # Lower-cased Latin letters can spell pinyin, so keep
+ # scripts apart unless the target mixes them: "AI" plus
+ # a character read "li" must not become a Chinese
+ # target read "ali", and a Latin target must not
+ # replace a Chinese character read "ai".
+ if (segment_latin and not target_latin) or (segment_han and not target_han):
+ continue
score = self._fuzz.ratio(segment_py, target_py) / 100.0
if score >= self.threshold:
candidates.append(
@@ -351,6 +379,25 @@ class PostprocessHotwordMatcher:
if not candidates:
return text, []
+ # A target already spelled correctly must not be rewritten into itself
+ # by a neighbouring window that only resembles it. Only candidates for
+ # the same target are dropped, so a short target found inside a longer
+ # one does not block the longer one.
+ exact_spans: Dict[str, List[Tuple[int, int]]] = {}
+ for target in {c.replacement for c in candidates}:
+ spans = exact_spans.setdefault(target, [])
+ idx = text.find(target)
+ while idx >= 0:
+ spans.append((idx, idx + len(target)))
+ idx = text.find(target, idx + 1)
+ candidates = [
+ c
+ for c in candidates
+ if not any(c.start < end and start < c.end for start, end in exact_spans[c.replacement])
+ ]
+ if not candidates:
+ return text, []
+
selected = _select_non_overlapping(candidates)
updated = text
applied: List[HotwordMatch] = []
diff --git a/tests/test_postprocess_hotwords.py b/tests/test_postprocess_hotwords.py
index d00a79086..330219d5e 100644
--- a/tests/test_postprocess_hotwords.py
+++ b/tests/test_postprocess_hotwords.py
@@ -79,6 +79,65 @@ class TestPostprocessHotwordMatcher(unittest.TestCase):
self.assertTrue(matches)
self.assertEqual(matches[0].replacement, "科大讯飞")
+ def test_fuzzy_leaves_surrounding_text_alone(self):
+ try:
+ import pypinyin # noqa: F401
+ import rapidfuzz # noqa: F401
+ except ImportError:
+ self.skipTest("pypinyin and rapidfuzz are required for fuzzy tests")
+
+ matcher = PostprocessHotwordMatcher(fuzzy_targets=["科大讯飞"])
+ # Already correct: nothing may change.
+ for text in ["我在科大讯飞工作", "去科大讯飞", "欢迎来到科大讯飞。", "今天科大讯飞,发布了新模型。"]:
+ self.assertEqual(matcher.apply_text(text), (text, []))
+ # Misrecognized: only the hotword changes, adjacent characters stay.
+ self.assertEqual(matcher.apply_text("欢迎来到科大迅飞。")[0], "欢迎来到科大讯飞。")
+ self.assertEqual(matcher.apply_text("科大迅飞2024年报")[0], "科大讯飞2024年报")
+
+ def test_fuzzy_short_target_does_not_block_longer_target(self):
+ try:
+ import pypinyin # noqa: F401
+ import rapidfuzz # noqa: F401
+ except ImportError:
+ self.skipTest("pypinyin and rapidfuzz are required for fuzzy tests")
+
+ matcher = PostprocessHotwordMatcher(fuzzy_targets=["讯飞", "科大讯飞"])
+ self.assertEqual(matcher.apply_text("可达讯飞发布")[0], "科大讯飞发布")
+
+ def test_fuzzy_latin_target_does_not_match_unrelated_text(self):
+ try:
+ import pypinyin # noqa: F401
+ import rapidfuzz # noqa: F401
+ except ImportError:
+ self.skipTest("pypinyin and rapidfuzz are required for fuzzy tests")
+
+ matcher = PostprocessHotwordMatcher(fuzzy_targets=["FunASR"])
+ text = "please open the model page today"
+ self.assertEqual(matcher.apply_text(text), (text, []))
+ # Windows must not split a Latin word or a number.
+ matcher = PostprocessHotwordMatcher(fuzzy_targets=["AI"])
+ self.assertEqual(matcher.apply_text("he said it"), ("he said it", []))
+ matcher = PostprocessHotwordMatcher(fuzzy_targets=["Win11"])
+ self.assertEqual(matcher.apply_text("Win10系统"), ("Win10系统", []))
+
+ def test_fuzzy_keeps_latin_and_chinese_apart(self):
+ try:
+ import pypinyin # noqa: F401
+ import rapidfuzz # noqa: F401
+ except ImportError:
+ self.skipTest("pypinyin and rapidfuzz are required for fuzzy tests")
+
+ # "AI" lower-cased is the pinyin "ai"; with the next character it would
+ # otherwise score 0.8571 against "ali".
+ matcher = PostprocessHotwordMatcher(fuzzy_targets=["阿里"])
+ self.assertEqual(matcher.apply_text("在AI里面"), ("在AI里面", []))
+ # A Latin target must not replace a Chinese character read "ai".
+ matcher = PostprocessHotwordMatcher(fuzzy_targets=["AI"])
+ self.assertEqual(matcher.apply_text("我爱你"), ("我爱你", []))
+ # A target that mixes scripts still corrects its Chinese homophone.
+ matcher = PostprocessHotwordMatcher(fuzzy_targets=["AI助手"])
+ self.assertEqual(matcher.apply_text("打开爱助手")[0], "打开AI助手")
+
def test_missing_fuzzy_dependency_raises(self):
with mock.patch.object(
postprocess_hotwords,