Adding fix for multi-byte segments in whisper.cpp (#734)

Co-authored-by: Chidi Williams <williamschidi1@gmail.com>

buzz/transcriber/whisper_cpp.py

@@ -15,6 +15,26 @@
 class WhisperCpp:
     def __init__(self, model: str) -> None:
         self.ctx = whisper_cpp.whisper_init_from_file(model.encode())
+        self.segments: List[Segment] = []
+
+    def append_segment(self, txt: bytes, start: int, end: int):
+        if txt == b'':
+            return True
+
+        # try-catch will guard against multi-byte utf-8 characters
+        # https://github.com/ggerganov/whisper.cpp/issues/1798
+        try:
+            self.segments.append(
+                Segment(
+                    start=start * 10,  # centisecond to ms
+                    end=end * 10,  # centisecond to ms
+                    text=txt.decode("utf-8"),
+                )
+            )
+
+            return True
+        except UnicodeDecodeError:
+            return False
 
     def transcribe(self, audio: Union[np.ndarray, str], params: Any):
         if isinstance(audio, str):
@@ -29,25 +49,50 @@ def transcribe(self, audio: Union[np.ndarray, str], params: Any):
         if result != 0:
             raise Exception(f"Error from whisper.cpp: {result}")
 
-        segments: List[Segment] = []
+        n_segments = whisper_cpp.whisper_full_n_segments(self.ctx)
 
-        n_segments = whisper_cpp.whisper_full_n_segments((self.ctx))
-        for i in range(n_segments):
-            txt = whisper_cpp.whisper_full_get_segment_text((self.ctx), i)
-            t0 = whisper_cpp.whisper_full_get_segment_t0((self.ctx), i)
-            t1 = whisper_cpp.whisper_full_get_segment_t1((self.ctx), i)
+        if params.token_timestamps:
+            # Will process word timestamps
+            txt_buffer = b''
+            txt_start = 0
+            txt_end = 0
 
-            segments.append(
-                Segment(
-                    start=t0 * 10,  # centisecond to ms
-                    end=t1 * 10,  # centisecond to ms
-                    text=txt.decode("utf-8"),
-                )
-            )
+            for i in range(n_segments):
+                txt = whisper_cpp.whisper_full_get_segment_text(self.ctx, i)
+                start = whisper_cpp.whisper_full_get_segment_t0(self.ctx, i)
+                end = whisper_cpp.whisper_full_get_segment_t1(self.ctx, i)
+
+                if txt.startswith(b' ') and self.append_segment(txt_buffer, txt_start, txt_end):
+                    txt_buffer = txt
+                    txt_start = start
+                    txt_end = end
+                    continue
+
+                if txt.startswith(b', '):
+                    txt_buffer += b','
+                    self.append_segment(txt_buffer, txt_start, txt_end)
+                    txt_buffer = txt.lstrip(b',')
+                    txt_start = start
+                    txt_end = end
+                    continue
+
+                txt_buffer += txt
+                txt_end = end
+
+            # Append the last segment
+            self.append_segment(txt_buffer, txt_start, txt_end)
+
+        else:
+            for i in range(n_segments):
+                txt = whisper_cpp.whisper_full_get_segment_text(self.ctx, i)
+                start = whisper_cpp.whisper_full_get_segment_t0(self.ctx, i)
+                end = whisper_cpp.whisper_full_get_segment_t1(self.ctx, i)
+
+                self.append_segment(txt, start, end)
 
         return {
-            "segments": segments,
-            "text": "".join([segment.text for segment in segments]),
+            "segments": self.segments,
+            "text": "".join([segment.text for segment in self.segments]),
         }
 
     def __del__(self):

tests/audio.py

@@ -3,3 +3,7 @@
 test_audio_path = os.path.abspath(
     os.path.join(os.path.dirname(__file__), "../testdata/whisper-french.mp3")
 )
+
+test_multibyte_utf8_audio_path = os.path.abspath(
+    os.path.join(os.path.dirname(__file__), "../testdata/whisper-latvian.wav")
+)

tests/transcriber/whisper_cpp_file_transcriber_test.py

@@ -13,7 +13,7 @@
     FileTranscriptionTask,
 )
 from buzz.transcriber.whisper_cpp_file_transcriber import WhisperCppFileTranscriber
-from tests.audio import test_audio_path
+from tests.audio import test_audio_path, test_multibyte_utf8_audio_path
 from tests.model_loader import get_model_path
 
 
@@ -25,7 +25,7 @@ class TestWhisperCppFileTranscriber:
                 False,
                 [Segment(0, 6560, "Bienvenue dans Passe-Relle. Un podcast pensé pour")],
             ),
-            (True, [Segment(30, 330, "Bien"), Segment(330, 740, "venue")]),
+            (True, [Segment(30, 740, "Bienvenue"), Segment(740, 1070, " dans")]),
         ],
     )
     def test_transcribe(
@@ -75,3 +75,63 @@ def test_transcribe(
             assert expected_segment.start == segments[i].start
             assert expected_segment.end == segments[i].end
             assert expected_segment.text in segments[i].text
+
+    @pytest.mark.parametrize(
+        "word_level_timings,expected_segments",
+        [
+            (
+                False,
+                [Segment(0, 7000, " Mani uzstrauts, laikabstākļi, tapēc uz jūru, es diezvajī braukša.")],
+            ),
+            (True, [Segment(380, 500, " Mani"), Segment(500, 1880, " uzstrauts,"), Segment(1880, 3920, " laikabstākļi")]),
+        ],
+    )
+    # Problematic part is in "laikabstākļi" where "ļ" gets returned from whisper.cpp in two segments
+    # First segment has first byte b'\xc4' and the second has second byte b'\xbc'.
+    def test_transcribe_latvian(
+        self, qtbot: QtBot, word_level_timings: bool, expected_segments: List[Segment]
+    ):
+        file_transcription_options = FileTranscriptionOptions(
+            file_paths=[test_multibyte_utf8_audio_path]
+        )
+        transcription_options = TranscriptionOptions(
+            language="lv",
+            task=Task.TRANSCRIBE,
+            word_level_timings=word_level_timings,
+            model=TranscriptionModel(
+                model_type=ModelType.WHISPER_CPP,
+                whisper_model_size=WhisperModelSize.TINY,
+            ),
+        )
+
+        model_path = get_model_path(transcription_options.model)
+        transcriber = WhisperCppFileTranscriber(
+            task=FileTranscriptionTask(
+                file_path=test_multibyte_utf8_audio_path,
+                transcription_options=transcription_options,
+                file_transcription_options=file_transcription_options,
+                model_path=model_path,
+            )
+        )
+        mock_progress = Mock(side_effect=lambda value: print("progress: ", value))
+        mock_completed = Mock()
+        mock_error = Mock()
+        transcriber.progress.connect(mock_progress)
+        transcriber.completed.connect(mock_completed)
+        transcriber.error.connect(mock_error)
+
+        with qtbot.wait_signal(transcriber.completed, timeout=10 * 60 * 1000):
+            transcriber.run()
+
+        mock_error.assert_not_called()
+
+        mock_progress.assert_called()
+        segments = [
+            segment
+            for segment in mock_completed.call_args[0][0]
+            if len(segment.text) > 0
+        ]
+        for i, expected_segment in enumerate(expected_segments):
+            assert expected_segment.start == segments[i].start
+            assert expected_segment.end == segments[i].end
+            assert expected_segment.text in segments[i].text