Merge branch 'master' of https://gitea.xiaowang.eu.org/wx44wx/AI-VideoAssistant

engine/core/duplex_pipeline.py

@@ -53,6 +53,7 @@ class DuplexPipeline:
     _SENTENCE_END_CHARS = frozenset({"。", "！", "？", ".", "!", "?", "\n"})
     _SENTENCE_TRAILING_CHARS = frozenset({"。", "！", "？", ".", "!", "?", "…", "~", "～", "\n"})
     _SENTENCE_CLOSERS = frozenset({'"', "'", "”", "’", ")", "]", "}", "）", "】", "」", "』", "》"})
+    _MIN_SPLIT_SPOKEN_CHARS = 6
 
     def __init__(
         self,
@@ -628,8 +629,11 @@ class DuplexPipeline:
         if not text_buffer:
             return None
 
+        search_start = 0
+        while True:
             split_idx = -1
-        for idx, char in enumerate(text_buffer):
+            for idx in range(search_start, len(text_buffer)):
+                char = text_buffer[idx]
                 if char == "." and self._is_non_sentence_period(text_buffer, idx):
                     continue
                 if char in self._SENTENCE_END_CHARS:
@@ -651,6 +655,17 @@ class DuplexPipeline:
                 return None
 
             sentence = text_buffer[:end_idx].strip()
+            spoken_chars = sum(1 for ch in sentence if ch.isalnum())
+
+            # Keep short utterances (e.g. "好。", "OK.") merged with following text.
+            if (
+                not force
+                and 0 < spoken_chars < self._MIN_SPLIT_SPOKEN_CHARS
+                and end_idx < len(text_buffer)
+            ):
+                search_start = end_idx
+                continue
+
             remainder = text_buffer[end_idx:]
             return sentence, remainder