fix: preserve transcripts without speaker tags

README.md

@@ -65,6 +65,8 @@ Timestamps are expressed in seconds, and adjacent segments are concatenated into
 [0.48][S01]Welcome everyone[1.66][12.26][S02]The new transcription pipeline is ready for evaluation[13.81][14.36][S01]Great, include the diarization results in the report[18.76]
 ```
 
+When a backend returns an otherwise valid segment without a speaker token, the local `parse_transcript` helper assigns the default speaker `S01` instead of dropping the segment. This only affects local parsing; the raw model text is preserved unchanged. Pass `default_speaker=None` when strict rejection of unlabelled segments is required.
+
 ## Model Architecture
 
 <p align="center">

README_zh.md

@@ -65,6 +65,8 @@ MOSS-Transcribe-Diarize 面向会议、通话、播客、访谈、讲座和视
 [0.48][S01]Welcome everyone[1.66][12.26][S02]The new transcription pipeline is ready for evaluation[13.81][14.36][S01]Great, include the diarization results in the report[18.76]
 ```
 
+当后端返回的片段格式有效、但缺少说话人标签时，本地 `parse_transcript` 辅助函数会使用默认说话人 `S01`，而不是丢弃整段内容。此行为只影响本地解析，原始模型文本会保持不变；如果需要严格拒绝缺少标签的片段，可传入 `default_speaker=None`。
+
 ## 模型架构
 
 <p align="center">

moss_transcribe_diarize/transcript_parser.py

@@ -23,6 +23,10 @@ class TranscriptStreamParser:
 
         [start][Sxx]text[end]
 
+    If the model omits the speaker token, ``default_speaker`` is used so that
+    otherwise valid transcript text is not discarded. Pass ``None`` to keep
+    strict parsing and reject unlabelled segments.
+
     The parser deliberately avoids regular expressions. It scans characters
     once, keeps only the active token/text buffers, and emits a segment after
     an end timestamp is confirmed by the next segment start or by ``close()``.
@@ -36,9 +40,20 @@ class TranscriptStreamParser:
     _READ_END = 5
     _AFTER_END = 6
 
-    def __init__(self, *, strip_text: bool = True, skip_empty: bool = True):
+    def __init__(
+        self,
+        *,
+        strip_text: bool = True,
+        skip_empty: bool = True,
+        default_speaker: str | None = "S01",
+    ):
+        if default_speaker is not None and _parse_speaker(list(default_speaker)) is None:
+            raise TranscriptParseError(
+                f"default_speaker must look like [Sxx], got {default_speaker!r}"
+            )
         self.strip_text = strip_text
         self.skip_empty = skip_empty
+        self.default_speaker = default_speaker
         self._state = self._SEEK_START
         self._token: list[str] = []
         self._text: list[str] = []
@@ -127,6 +142,11 @@ def _expect_speaker_open(self, ch: str) -> None:
         if ch == "[":
             self._token.clear()
             self._state = self._READ_SPEAKER
+        elif not ch.isspace() and self.default_speaker is not None:
+            self._speaker = self.default_speaker
+            self._text.clear()
+            self._state = self._READ_TEXT
+            self._read_text(ch)
         elif not ch.isspace():
             self.reset()
 

tests/test_subtitle_postprocess.py

@@ -18,6 +18,20 @@ def test_builds_subtitle_segments_from_transcript(self):
         self.assertEqual([segment.text for segment in segments], ["你好", "开始"])
         self.assertEqual(segments[0].id, "seg_0001")
 
+    def test_recovers_unlabelled_single_speaker_segments(self):
+        segments = subtitle_segments_from_transcript(
+            "[0.0]連續單一說話者內容[2.0][2.0]下一段[3.0]",
+            postprocess=False,
+        )
+
+        self.assertEqual(
+            [(segment.start, segment.end, segment.speaker, segment.text) for segment in segments],
+            [
+                (0.0, 2.0, "S01", "連續單一說話者內容"),
+                (2.0, 3.0, "S01", "下一段"),
+            ],
+        )
+
     def test_can_build_raw_subtitle_segments_without_postprocess(self):
         segments = subtitle_segments_from_transcript(
             "[0][S01]短[0.4][0.2][S01]重叠但保留[0.8]",

tests/test_transcript_parser.py

@@ -3,6 +3,7 @@
 import unittest
 
 from moss_transcribe_diarize.transcript_parser import (
+    TranscriptParseError,
     TranscriptSegment,
     TranscriptStreamParser,
     iter_transcript_segments,
@@ -58,6 +59,35 @@ def test_numeric_brackets_inside_text_are_preserved(self):
             [TranscriptSegment(0.0, 4.0, "S01", "第[2024]年，编号[001]继续")],
         )
 
+    def test_missing_speaker_uses_default_speaker(self):
+        text = "[0.0]單一說話者片段[2.0][2.0]下一段[3.0]"
+
+        self.assertEqual(
+            parse_transcript(text),
+            [
+                TranscriptSegment(0.0, 2.0, "S01", "單一說話者片段"),
+                TranscriptSegment(2.0, 3.0, "S01", "下一段"),
+            ],
+        )
+
+    def test_missing_speaker_can_be_parsed_from_arbitrary_chunks(self):
+        chunks = ["[0.0]hel", "lo[1.0]", "[1.0]world[2.0]"]
+
+        self.assertEqual(
+            list(iter_transcript_segments(chunks)),
+            [
+                TranscriptSegment(0.0, 1.0, "S01", "hello"),
+                TranscriptSegment(1.0, 2.0, "S01", "world"),
+            ],
+        )
+
+    def test_strict_mode_keeps_unlabelled_segments_ignored(self):
+        self.assertEqual(parse_transcript("[0]unlabelled[1]", default_speaker=None), [])
+
+    def test_default_speaker_must_be_valid(self):
+        with self.assertRaises(TranscriptParseError):
+            TranscriptStreamParser(default_speaker="speaker-1")
+
     def test_noise_before_first_segment_is_ignored(self):
         text = "noise [bad][0.1][S01]hello[0.9]"