optimized PC player

sentence_api/transcription.py

@@ -246,6 +246,59 @@ _ABBREVIATION_RE = re.compile(
     r"e\.g\.|i\.e\.|no\.|approx\.|fig\.|inc\.|ltd\.|jr\.|sr\.|u\.s\.|u\.k\.)"
     r"(?![A-Za-z0-9])"
 )
+_SPEAKER_TAG_RE = re.compile(r"\[(S\d+)\]")
+
+
+def _expand_speaker_segments(
+    segment: TranscriptionSegment,
+) -> List[TranscriptionSegment]:
+    """Split a segment at inline ``[S01]`` speaker tags and strip the tags.
+
+    MOSS sometimes embeds speaker tags directly in the text (rather than in
+    the ``speaker`` JSON field). Each tag marks the start of a new speaker
+    turn, so the text is split there and the tags are removed from the
+    visible transcription.
+    """
+    text = segment.text
+    matches = list(_SPEAKER_TAG_RE.finditer(text))
+    if not matches:
+        return [segment]
+
+    sub_segments: List[TranscriptionSegment] = []
+    total = len(matches)
+    duration = segment.end_seconds - segment.start_seconds
+    for index, match in enumerate(matches):
+        start_pos = match.end()
+        if index + 1 < total:
+            end_pos = matches[index + 1].start()
+        else:
+            end_pos = len(text)
+        piece = text[start_pos:end_pos].strip()
+        if not piece:
+            continue
+        # Allocate time proportionally across sub-segments.
+        frac_start = index / total
+        frac_end = (index + 1) / total
+        sub_start = segment.start_seconds + duration * frac_start
+        sub_end = segment.start_seconds + duration * frac_end
+        words = None
+        if segment.words:
+            sub_words = [
+                w
+                for w in segment.words
+                if w.start_seconds >= sub_start and w.end_seconds <= sub_end
+            ]
+            words = sub_words or None
+        sub_segments.append(
+            TranscriptionSegment(
+                start_seconds=sub_start,
+                end_seconds=sub_end,
+                text=piece,
+                speaker=match.group(1),
+                words=words,
+            )
+        )
+    return sub_segments or [segment]
 
 
 def split_sentences_at_punctuation(
@@ -257,14 +310,18 @@ def split_sentences_at_punctuation(
     across segments and a sentence is closed at periods/question marks. Commas
     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.
     """
     sentences: List[TranscriptionSegment] = []
     pieces: List[str] = []
     start_seconds: Optional[float] = None
     last_segment_end: Optional[float] = None
+    current_speaker: Optional[str] = None
 
     def flush(end_seconds: float) -> None:
-        nonlocal pieces, start_seconds
+        nonlocal pieces, start_seconds, current_speaker
         text = " ".join(piece for piece in pieces if piece).strip()
         if text and start_seconds is not None and end_seconds > start_seconds:
             sentences.append(
@@ -272,19 +329,31 @@ def split_sentences_at_punctuation(
                     start_seconds=start_seconds,
                     end_seconds=end_seconds,
                     text=text,
+                    speaker=current_speaker,
                 )
             )
         pieces = []
         start_seconds = None
+        current_speaker = None
+
+    expanded: List[TranscriptionSegment] = []
+    for seg in segments:
+        expanded.extend(_expand_speaker_segments(seg))
 
     for segment in sorted(
-        segments, key=lambda item: (item.start_seconds, item.end_seconds)
+        expanded, key=lambda item: (item.start_seconds, item.end_seconds)
     ):
         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:
+            flush(segment.start_seconds)
         if not pieces:
             start_seconds = segment.start_seconds
+        current_speaker = segment.speaker
         seg_pieces: List[Tuple[str, Optional[str]]] = []
         cursor = 0
         for match in _PUNCTUATION_BOUNDARY.finditer(text):
@@ -322,6 +391,7 @@ def split_sentences_at_punctuation(
                 end_seconds = ends[index] if index < len(ends) else segment.end_seconds
                 flush(end_seconds)
                 start_seconds = end_seconds
+                current_speaker = segment.speaker
     if pieces:
         flush(last_segment_end if last_segment_end is not None else start_seconds)
     return sentences