fixed a bug

sentence_api/DEPLOYMENT.md

@@ -152,10 +152,11 @@ curl -X POST http://127.0.0.1:8001/v1/audio/transcriptions \
 
 ### 断句规则
 
-`sentence_api` 不以 Whisper 返回的 segment 直接作为句子，而是把每个 segment 的文本
+`sentence_api` 不以 Whisper 返回的 segment 直接作为句子：文本会跨 segment 累积，
-按句号（`.`）拆成一句句，句号即句子结束。若转写服务支持词级时间戳
+只有碰到句号（`.`）或问号（`?`）才算一句结束；Whisper 在没有标点处断开的
+segment 会自动合并成同一句。若转写服务支持词级时间戳
 （`timestamp_granularities[]=word`，Whisper/Speaches 支持），句子的结束时间用
-句号所在单词的时间戳；不支持时自动按文字长度在 segment 内估算，请求失败会回退为
+标点所在单词的时间戳；不支持时自动按文字长度在 segment 内估算，请求失败会回退为
 不带词级时间戳的普通请求。MOSS 不支持该参数时同样自动回退，不影响转写。
 
 ## 3. 部署 API

sentence_api/processing.py

@@ -9,7 +9,7 @@ from .config import Settings
 from .generate_boundaries import ALGORITHM_VERSION, make_entry
 from .models import SentenceBoundary, SentenceBoundaryDocument
 from .repository import VideoRepository
-from .transcription import Transcript, Transcriber, split_segment_by_periods
+from .transcription import Transcript, Transcriber, split_sentences_at_punctuation
 
 
 MOSS_ALGORITHM_VERSION = "moss-period-v2"
@@ -107,30 +107,29 @@ def document_from_transcript(
     samples, sample_rate = decode_audio_mono(audio_path)
     sentences: List[SentenceBoundary] = []
     previous_end = 0
-    for segment in sorted(transcript.segments, key=lambda item: (item.start_seconds, item.end_seconds)):
+    for sentence_segment in split_sentences_at_punctuation(transcript.segments):
-        for sentence_segment in split_segment_by_periods(segment):
+        start_ms = max(previous_end, int(round(sentence_segment.start_seconds * 1000)))
-            start_ms = max(previous_end, int(round(sentence_segment.start_seconds * 1000)))
+        end_ms = min(duration_ms, int(round(sentence_segment.end_seconds * 1000)))
-            end_ms = min(duration_ms, int(round(sentence_segment.end_seconds * 1000)))
+        if not sentence_segment.text.strip() or end_ms <= start_ms:
-            if not sentence_segment.text.strip() or end_ms <= start_ms:
+            continue
-                continue
+        start_sample = max(0, int(start_ms / 1000 * sample_rate))
-            start_sample = max(0, int(start_ms / 1000 * sample_rate))
+        end_sample = min(samples.size, int(end_ms / 1000 * sample_rate))
-            end_sample = min(samples.size, int(end_ms / 1000 * sample_rate))
+        try:
-            try:
+            metrics = analyze_samples(samples[start_sample:end_sample], sample_rate)
-                metrics = analyze_samples(samples[start_sample:end_sample], sample_rate)
+            speech_duration_ms = metrics.speech_duration_ms
-                speech_duration_ms = metrics.speech_duration_ms
+        except AudioAnalysisError:
-            except AudioAnalysisError:
+            speech_duration_ms = end_ms - start_ms
-                speech_duration_ms = end_ms - start_ms
+        sentences.append(
-            sentences.append(
+            SentenceBoundary(
-                SentenceBoundary(
+                index=len(sentences),
-                    index=len(sentences),
+                start_ms=start_ms,
-                    start_ms=start_ms,
+                end_ms=end_ms,
-                    end_ms=end_ms,
+                text=sentence_segment.text.strip(),
-                    text=sentence_segment.text.strip(),
+                language=language,
-                    language=language,
+                reference_speech_duration_ms=max(1, speech_duration_ms),
-                    reference_speech_duration_ms=max(1, speech_duration_ms),
-                )
             )
-            previous_end = end_ms
+        )
+        previous_end = end_ms
     return SentenceBoundaryDocument(
         video_hash=video_hash,
         duration_ms=duration_ms,

sentence_api/tests/test_processing.py

@@ -22,48 +22,61 @@ def _silence(seconds: float) -> np.ndarray:
     return np.zeros(int(seconds * SAMPLE_RATE))
 
 
-def test_document_from_transcript_splits_sentences_at_periods(tmp_path):
+def test_document_from_transcript_splits_sentences_at_punctuation(tmp_path):
     samples = np.concatenate(
         [
-            _speech(2.0),
+            _speech(8.0),
-            _silence(0.2),
+            _speech(1.5),
-            _speech(2.0),
+            _speech(10.5),
-            _silence(0.2),
+            _speech(4.0),
-            _speech(1.6),
         ]
     ).astype(np.float32)
     wav = tmp_path / "audio.wav"
     wav.write_bytes(_wav_bytes(samples, SAMPLE_RATE))
     transcript = Transcript(
-        text="Hello world. Good day everyone. Nice to meet you.",
+        text=(
+            "So I've been single for about four years now and I find it hard to meet "
+            "a guy especially in California. I've tried dating apps, I've met friends "
+            "through friends, I've done online dating and I just feel like most of the "
+            "guys that I meet aren't serious about a relationship."
+        ),
         segments=[
             TranscriptionSegment(
                 start_seconds=0.0,
-                end_seconds=2.0,
+                end_seconds=8.0,
-                text="Hello world.",
+                text=(
-                words=[
+                    "So I've been single for about four years now and I find it hard "
-                    WordTimestamp(0.0, 0.6, "Hello"),
+                    "to meet a guy especially in"
-                    WordTimestamp(0.7, 1.5, "world."),
+                ),
-                ],
             ),
             TranscriptionSegment(
-                start_seconds=2.2,
+                start_seconds=8.0,
-                end_seconds=5.5,
+                end_seconds=9.5,
-                text="Good day everyone. Nice to meet you.",
+                text="California.",
-                words=[
+                words=[WordTimestamp(8.0, 9.5, "California.")],
-                    WordTimestamp(2.2, 2.8, "Good"),
+            ),
-                    WordTimestamp(2.9, 3.5, "day"),
+            TranscriptionSegment(
-                    WordTimestamp(3.6, 4.2, "everyone."),
+                start_seconds=9.5,
-                    WordTimestamp(4.4, 4.9, "Nice"),
+                end_seconds=20.0,
-                    WordTimestamp(5.0, 5.5, "you."),
+                text=(
-                ],
+                    "I've tried dating apps, I've met friends through friends, "
+                    "I've done online dating"
+                ),
+            ),
+            TranscriptionSegment(
+                start_seconds=20.0,
+                end_seconds=24.0,
+                text=(
+                    "and I just feel like most of the guys that I meet aren't serious "
+                    "about a relationship."
+                ),
             ),
         ],
     )
 
     document = document_from_transcript(
         video_hash=VIDEO_HASH,
-        duration_ms=6000,
+        duration_ms=24000,
         transcript=transcript,
         language="en",
         audio_path=wav,
@@ -71,14 +84,18 @@ def test_document_from_transcript_splits_sentences_at_periods(tmp_path):
 
     assert document.algorithm_version == MOSS_ALGORITHM_VERSION
     assert [sentence.text for sentence in document.sentences] == [
-        "Hello world.",
+        (
-        "Good day everyone.",
+            "So I've been single for about four years now and I find it hard to meet "
-        "Nice to meet you.",
+            "a guy especially in California."
+        ),
+        (
+            "I've tried dating apps, I've met friends through friends, I've done "
+            "online dating and I just feel like most of the guys that I meet aren't "
+            "serious about a relationship."
+        ),
     ]
     assert document.sentences[0].start_ms == 0
-    assert document.sentences[0].end_ms == 1500
+    assert document.sentences[0].end_ms == 9500
-    assert document.sentences[1].start_ms == 2200
+    assert document.sentences[1].start_ms == 9500
-    assert document.sentences[1].end_ms == 4200
+    assert document.sentences[1].end_ms == 24000
-    assert document.sentences[2].start_ms == 4200
-    assert document.sentences[2].end_ms == 5500
     assert all(sentence.reference_speech_duration_ms > 0 for sentence in document.sentences)

sentence_api/tests/test_transcription.py

@@ -11,7 +11,7 @@ from sentence_api.transcription import (
     _parse_json_segments,
     _plan_chunks,
     _parse_word_timestamps,
-    split_segment_by_periods,
+    split_sentences_at_punctuation,
     _wav_bytes,
     _wav_duration_seconds,
 )
@@ -195,7 +195,59 @@ def test_parse_word_timestamps_ignores_invalid_entries():
     assert _parse_word_timestamps("nope") is None
 
 
-def test_split_segment_by_periods_uses_word_timestamps():
+def test_split_sentences_merges_segments_without_punctuation():
+    segments = [
+        TranscriptionSegment(
+            start_seconds=0.0,
+            end_seconds=8.0,
+            text=(
+                "So I've been single for about four years now and I find it hard "
+                "to meet a guy especially in"
+            ),
+        ),
+        TranscriptionSegment(
+            start_seconds=8.0,
+            end_seconds=9.5,
+            text="California.",
+            words=[WordTimestamp(8.0, 9.5, "California.")],
+        ),
+        TranscriptionSegment(
+            start_seconds=9.5,
+            end_seconds=20.0,
+            text=(
+                "I've tried dating apps, I've met friends through friends, "
+                "I've done online dating"
+            ),
+        ),
+        TranscriptionSegment(
+            start_seconds=20.0,
+            end_seconds=24.0,
+            text=(
+                "and I just feel like most of the guys that I meet aren't serious "
+                "about a relationship."
+            ),
+            words=[WordTimestamp(20.0, 24.0, "relationship.")],
+        ),
+    ]
+    sentences = split_sentences_at_punctuation(segments)
+    assert [(s.text, s.start_seconds, s.end_seconds) for s in sentences] == [
+        (
+            "So I've been single for about four years now and I find it hard to meet "
+            "a guy especially in California.",
+            0.0,
+            9.5,
+        ),
+        (
+            "I've tried dating apps, I've met friends through friends, I've done "
+            "online dating and I just feel like most of the guys that I meet aren't "
+            "serious about a relationship.",
+            9.5,
+            24.0,
+        ),
+    ]
+
+
+def test_split_sentences_uses_word_timestamps_within_segment():
     segment = TranscriptionSegment(
         start_seconds=0.0,
         end_seconds=4.0,
@@ -209,7 +261,7 @@ def test_split_segment_by_periods_uses_word_timestamps():
             WordTimestamp(3.9, 4.0, "you."),
         ],
     )
-    sentences = split_segment_by_periods(segment)
+    sentences = split_sentences_at_punctuation([segment])
     assert [(s.text, s.start_seconds, s.end_seconds) for s in sentences] == [
         ("Hello world.", 0.0, 1.5),
         ("Good day.", 1.5, 3.2),
@@ -217,27 +269,46 @@ def test_split_segment_by_periods_uses_word_timestamps():
     ]
 
 
-def test_split_segment_by_periods_falls_back_to_proportional():
+def test_split_sentences_splits_at_question_marks():
     segment = TranscriptionSegment(
-        start_seconds=10.0,
+        start_seconds=0.0,
-        end_seconds=20.0,
+        end_seconds=10.0,
-        text="First sentence. Second sentence. Third.",
+        text="Really? Wait a minute. What about now?",
     )
-    sentences = split_segment_by_periods(segment)
+    sentences = split_sentences_at_punctuation([segment])
     assert [s.text for s in sentences] == [
-        "First sentence.",
+        "Really?",
-        "Second sentence.",
+        "Wait a minute.",
-        "Third.",
+        "What about now?",
+    ]
+    assert sentences[0].start_seconds == 0.0
+    assert sentences[0].end_seconds == sentences[1].start_seconds
+    assert sentences[1].end_seconds == sentences[2].start_seconds
+    assert sentences[2].end_seconds == 10.0
+
+
+def test_split_sentences_falls_back_to_proportional():
+    segments = [
+        TranscriptionSegment(
+            start_seconds=10.0, end_seconds=14.0, text="This is the first"
+        ),
+        TranscriptionSegment(
+            start_seconds=14.0,
+            end_seconds=18.0,
+            text="sentence. And this is the second one.",
+        ),
+    ]
+    sentences = split_sentences_at_punctuation(segments)
+    assert [s.text for s in sentences] == [
+        "This is the first sentence.",
+        "And this is the second one.",
     ]
     assert sentences[0].start_seconds == 10.0
-    assert sentences[1].start_seconds == sentences[0].end_seconds
+    assert sentences[0].end_seconds == sentences[1].start_seconds
-    assert sentences[2].end_seconds == 20.0
+    assert sentences[1].end_seconds == 18.0
-    assert sentences[0].end_seconds > 10.0
-    assert sentences[1].end_seconds < 20.0
-    assert sentences[0].end_seconds < sentences[1].end_seconds
 
 
-def test_split_segment_by_periods_falls_back_when_word_ends_are_invalid():
+def test_split_sentences_falls_back_when_word_ends_are_invalid():
     segment = TranscriptionSegment(
         start_seconds=0.0,
         end_seconds=2.0,
@@ -247,30 +318,37 @@ def test_split_segment_by_periods_falls_back_when_word_ends_are_invalid():
             WordTimestamp(2.6, 3.0, "Two."),
         ],
     )
-    sentences = split_segment_by_periods(segment)
+    sentences = split_sentences_at_punctuation([segment])
     assert [s.text for s in sentences] == ["One.", "Two."]
     assert sentences[0].end_seconds == sentences[1].start_seconds
     assert sentences[1].end_seconds == 2.0
 
 
-def test_split_segment_by_periods_keeps_segment_without_period():
+def test_split_sentences_keeps_incomplete_tail_as_sentence():
     segment = TranscriptionSegment(
         start_seconds=1.0, end_seconds=2.0, text="no period here"
     )
-    assert split_segment_by_periods(segment) == [segment]
+    sentences = split_sentences_at_punctuation([segment])
+    assert [(s.text, s.start_seconds, s.end_seconds) for s in sentences] == [
+        ("no period here", 1.0, 2.0)
+    ]
+    assert split_sentences_at_punctuation([]) == []
 
 
-def test_split_segment_by_periods_handles_ellipsis_and_dots_only():
+def test_split_sentences_handles_ellipsis_and_dots_only():
     segment = TranscriptionSegment(
         start_seconds=0.0,
         end_seconds=2.0,
         text="Wait... What? ...",
     )
-    sentences = split_segment_by_periods(segment)
+    sentences = split_sentences_at_punctuation([segment])
-    assert [s.text for s in sentences] == ["Wait...", "What? ..."]
+    assert [s.text for s in sentences] == ["Wait...", "What?"]
     assert sentences[0].start_seconds == 0.0
     assert sentences[0].end_seconds == sentences[1].start_seconds
     assert sentences[1].end_seconds == 2.0
+    assert split_sentences_at_punctuation(
+        [TranscriptionSegment(0.0, 1.0, "...")]
+    ) == []
 
 
 def test_post_audio_requests_word_timestamps_and_falls_back(monkeypatch):

sentence_api/transcription.py

@@ -239,55 +239,75 @@ def _parse_word_timestamps(raw_words) -> Optional[List[WordTimestamp]]:
     return words or None
 
 
-_PERIOD_BOUNDARY = re.compile(r"\.+")
+_PUNCTUATION_BOUNDARY = re.compile(r"[.?]+")
 
 
-def split_segment_by_periods(
+def split_sentences_at_punctuation(
-    segment: TranscriptionSegment,
+    segments: List[TranscriptionSegment],
 ) -> List[TranscriptionSegment]:
-    """Split one whisper segment into sentences at periods (".")."""
+    """Merge whisper segments and cut sentences at periods/question marks.
-    text = segment.text.strip()
+
-    if not text:
+    Whisper's own segment breaks often fall mid-sentence, so text is accumulated
-        return []
+    across segments and a sentence is only closed once the accumulated text
+    reaches a period or a question mark.
+    """
+    sentences: List[TranscriptionSegment] = []
     pieces: List[str] = []
-    cursor = 0
+    start_seconds: Optional[float] = None
-    for match in _PERIOD_BOUNDARY.finditer(text):
+    last_segment_end: Optional[float] = None
-        piece = text[cursor:match.end()].strip()
+
-        if piece.strip(".").strip():
+    def flush(end_seconds: float) -> None:
-            pieces.append(piece)
+        nonlocal pieces, start_seconds
-        cursor = match.end()
+        text = " ".join(piece for piece in pieces if piece).strip()
-    tail = text[cursor:].strip()
+        if text and start_seconds is not None and end_seconds > start_seconds:
-    if tail:
+            sentences.append(
-        pieces.append(tail)
-    if len(pieces) <= 1:
-        return [segment] if pieces else []
-    ends = _sentence_end_seconds(segment, pieces)
-    sub_segments: List[TranscriptionSegment] = []
-    start = segment.start_seconds
-    for piece, end in zip(pieces, ends):
-        if end > start:
-            sub_segments.append(
                 TranscriptionSegment(
-                    start_seconds=start,
+                    start_seconds=start_seconds,
-                    end_seconds=end,
+                    end_seconds=end_seconds,
-                    text=piece,
+                    text=text,
-                    speaker=segment.speaker,
                 )
             )
-        start = end
+        pieces = []
-    return sub_segments
+        start_seconds = None
 
-
+    for segment in sorted(
-def _sentence_end_seconds(
+        segments, key=lambda item: (item.start_seconds, item.end_seconds)
-    segment: TranscriptionSegment, pieces: List[str]
+    ):
-) -> List[float]:
+        text = segment.text.strip()
-    lengths = [len(piece) for piece in pieces]
+        if not text:
-    total = sum(lengths)
+            continue
-    if segment.words and len(segment.words) >= len(pieces):
+        if not pieces:
-        word_ends = _word_boundary_ends(segment, lengths)
+            start_seconds = segment.start_seconds
-        if _valid_boundaries(segment, word_ends):
+        seg_pieces: List[str] = []
-            return word_ends + [segment.end_seconds]
+        cursor = 0
-    return _proportional_ends(segment, lengths) + [segment.end_seconds]
+        for match in _PUNCTUATION_BOUNDARY.finditer(text):
+            piece = text[cursor:match.end()].strip()
+            cursor = match.end()
+            if piece.strip(".?").strip():
+                seg_pieces.append(piece)
+        tail = text[cursor:].strip()
+        if tail:
+            seg_pieces.append(tail)
+        if not seg_pieces:
+            continue
+        lengths = [len(piece) for piece in seg_pieces]
+        ends: List[float] = []
+        if segment.words and len(segment.words) >= len(seg_pieces):
+            ends = _word_boundary_ends(segment, lengths)
+            if not _valid_boundaries(segment, ends):
+                ends = []
+        if not ends:
+            ends = _proportional_ends(segment, lengths)
+        last_segment_end = segment.end_seconds
+        for index, piece in enumerate(seg_pieces):
+            pieces.append(piece)
+            if piece[-1] in ".?":
+                end_seconds = ends[index] if index < len(ends) else segment.end_seconds
+                flush(end_seconds)
+                start_seconds = end_seconds
+    if pieces:
+        flush(last_segment_end if last_segment_end is not None else start_seconds)
+    return sentences
 
 
 def _word_boundary_ends(