fixed a bug

sentence_api/transcription.py

@@ -246,9 +246,18 @@ _PUNCTUATION_BOUNDARY = re.compile(r"[.?]+|,+")
 MIN_WORDS_FOR_COMMA_SPLIT = 50
 _ABBREVIATION_RE = re.compile(
     r"(?i)\b(?:a\.m\.|p\.m\.|mr\.|mrs\.|ms\.|dr\.|prof\.|st\.|vs\.|etc\.|"
-    r"e\.g\.|i\.e\.|no\.|approx\.|fig\.|inc\.|ltd\.|jr\.|sr\.|u\.s\.|u\.k\.)"
+    r"e\.g\.|i\.e\.|no\.|approx\.|fig\.|inc\.|ltd\.|jr\.|sr\.|u\.s\.|u\.k\.|"
+    r"ph\.d\.|m\.d\.|b\.a\.|m\.a\.|d\.c\.|p\.s\.)"
     r"(?![A-Za-z0-9])"
 )
+_TIME_MERIDIEM_RE = re.compile(
+    r"(?i)\b(?:[01]?\d|2[0-3])(?::[0-5]\d)?\s*(?:a\.m\.|p\.m\.|am\.|pm\.)"
+    r"(?![A-Za-z0-9])"
+)
+_TIME_CONTINUATION_RE = re.compile(
+    r"(?i)^(eastern|western|central|pacific|mountain|time|"
+    r"et|est|edt|pt|pst|pdt|ct|cst|cdt|mt|mst|mdt|gmt|utc|local)\b"
+)
 _SPEAKER_TAG_RE = re.compile(r"\[(S\d+)\]")
 
 
@@ -319,8 +328,10 @@ def split_sentences_at_punctuation(
     are also eligible boundaries, but only when the text before the comma has
     enough words; otherwise the comma stays with the following text.
 
-    Inline ``[S01]`` speaker tags are stripped and each speaker turn always
-    starts a new sentence, even when no sentence-ending punctuation exists.
+    Inline ``[S01]`` speaker tags are stripped. A new sentence starts when the
+    speaker actually changes; same-speaker line wraps stay merged until
+    sentence punctuation. Decimal points, initials, and common abbreviations
+    are not treated as sentence ends.
     """
     sentences: List[TranscriptionSegment] = []
     pieces: List[str] = []
@@ -354,10 +365,14 @@ def split_sentences_at_punctuation(
         text = segment.text.strip()
         if not text:
             continue
-        # Each speaker-tagged sub-segment always starts a new sentence,
-        # regardless of whether the speaker actually changed — MOSS emits
-        # a tag per phrase/line even for the same speaker.
-        if pieces and segment.speaker is not None:
+        # MOSS often tags every phrase/line with [S01] even when the speaker
+        # did not change. Only a real speaker change starts a new sentence.
+        if (
+            pieces
+            and segment.speaker is not None
+            and current_speaker is not None
+            and segment.speaker != current_speaker
+        ):
             flush(segment.start_seconds)
         if not pieces:
             start_seconds = segment.start_seconds
@@ -418,31 +433,70 @@ def _is_sentence_end_run(text: str, run_start: int, run_end: int) -> bool:
     punctuation = text[run_start:run_end]
     if "?" in punctuation:
         return True
-    if (
-        run_start > 0
-        and run_end < len(text)
-        and text[run_start - 1].isdigit()
-        and text[run_end].isdigit()
-    ):
-        return False  # decimal point, e.g. "0.4"
+    if _is_decimal_point(text, run_start, run_end):
+        return False
     if len(punctuation) > 1:
         return True  # ellipsis or "!?" etc.
     if _is_abbreviation_run(text, run_start, run_end):
         return False  # e.g. "a.m.", "p.m.", "Mr."
+    if _is_time_meridiem_run(text, run_start, run_end):
+        return False  # e.g. "3:00 PM.", "5 pm."
+    if _is_single_letter_initial(text, run_start, run_end):
+        return False  # e.g. "Donald J. Trump"
     rest = text[run_end:].lstrip()
     if rest and rest[0].islower():
         return False  # period followed by a lowercase word is an abbreviation
     return True
 
 
+def _is_decimal_point(text: str, run_start: int, run_end: int) -> bool:
+    if run_end - run_start != 1 or text[run_start] != ".":
+        return False
+    if run_start == 0 or run_end >= len(text):
+        return False
+    return text[run_start - 1].isdigit() and text[run_end].isdigit()
+
+
+def _is_single_letter_initial(text: str, run_start: int, run_end: int) -> bool:
+    if run_end - run_start != 1 or text[run_start] != ".":
+        return False
+    if run_start == 0 or not text[run_start - 1].isalpha():
+        return False
+    return run_start == 1 or not text[run_start - 2].isalpha()
+
+
+def _meridiem_continues(text: str, after: int) -> bool:
+    rest = text[after:].lstrip()
+    if not rest:
+        return True
+    if rest[0].islower():
+        return True
+    return _TIME_CONTINUATION_RE.match(rest) is not None
+
+
+def _is_time_meridiem_run(text: str, run_start: int, run_end: int) -> bool:
+    for match in _TIME_MERIDIEM_RE.finditer(text):
+        if not (match.start() < run_end <= match.end()):
+            continue
+        if run_end < match.end():
+            return True
+        return _meridiem_continues(text, match.end())
+    return False
+
+
 def _is_abbreviation_run(text: str, run_start: int, run_end: int) -> bool:
     for match in _ABBREVIATION_RE.finditer(text):
         if not (match.start() < run_end <= match.end()):
             continue
-        if match.group(0).lower() == "no.":
+        token = match.group(0).lower()
+        if token == "no.":
             rest = text[match.end():].lstrip()
             if not rest or not rest[0].isdigit():
                 return False  # "No." as a reply ends the sentence
+        if token in {"a.m.", "p.m."}:
+            if run_end < match.end():
+                return True
+            return _meridiem_continues(text, match.end())
         return True
     return False