fix(sensevoice): keep words separate at standalone boundaries

Signed-off-by: LauraGPT <[email protected]>

funasr/models/sense_voice/model.py

@@ -1093,6 +1093,7 @@ class SenseVoiceSmall(nn.Module):
             start = int(start * 1000)
             end = int(end * 1000)
             if word == "▁":
+                prev_word = None
                 continue
             if i == 0:
                 word = word[1:] if word.startswith("▁") else word

tests/test_sensevoice_word_timestamps.py

@@ -66,6 +66,24 @@ from funasr.models.sense_voice.model import SenseVoiceSmall
         pytest.param(
             ["A", "\u2581B", "C"], ["A", "BC"], [[0, 125], [125, 375]], id="word-boundary"
         ),
+        pytest.param(
+            ["Hello", "\u2581", "world"],
+            ["Hello", "world"],
+            [[0, 125], [250, 375]],
+            id="standalone-word-boundary",
+        ),
+        pytest.param(
+            ["\u2581Hel", "lo", "\u2581", "wor", "ld"],
+            ["Hello", "world"],
+            [[0, 250], [375, 625]],
+            id="subpieces-around-standalone-boundary",
+        ),
+        pytest.param(
+            ["Hello", "\u2581", "\u2581", "world"],
+            ["Hello", "world"],
+            [[0, 125], [375, 500]],
+            id="repeated-standalone-boundaries",
+        ),
     ],
 )
 def test_post_preserves_word_boundaries_and_timestamp_alignment(tokens, words, timestamps):