xfyun asr update one bubble

engine/xfyun_asr.py

@@ -78,6 +78,12 @@ class XfyunASRService(STTService):
         self._sent_final_frame = False
         self._partials: list[str] = []
         self._last_text = ""
+        # Text already finalized by xfyun within the current VAD turn. xfyun
+        # may emit several status=2 segments within one turn (e.g. when the
+        # user pauses briefly); each segment resets `_partials`/`_last_text`,
+        # but interim frames pushed to clients should still grow
+        # monotonically across segments. Reset on VADUserStartedSpeakingFrame.
+        self._turn_committed_text = ""
 
     async def cleanup(self) -> None:
         await self._close_utterance()
@@ -95,6 +101,7 @@ class XfyunASRService(STTService):
         await super().process_frame(frame, direction)
 
         if isinstance(frame, VADUserStartedSpeakingFrame):
+            self._turn_committed_text = ""
             await self._start_utterance()
         elif isinstance(frame, VADUserStoppedSpeakingFrame):
             await self._finish_utterance()
@@ -261,7 +268,7 @@ class XfyunASRService(STTService):
                 self._last_text = text
                 await self.push_frame(
                     InterimTranscriptionFrame(
-                        text,
+                        self._turn_committed_text + text,
                         self._user_id,
                         time_now_iso8601(),
                         _language_or_none(self._language),
@@ -273,6 +280,10 @@ class XfyunASRService(STTService):
             final_text = self._last_text
             if final_text:
                 self.confirm_finalize()
+                # Emit just this segment's text. The pipecat user aggregator
+                # concatenates TranscriptionFrames within a VAD turn, so we
+                # must NOT prepend `_turn_committed_text` here or the
+                # aggregated turn text would double-count earlier segments.
                 await self.push_frame(
                     TranscriptionFrame(
                         final_text,
@@ -282,6 +293,9 @@ class XfyunASRService(STTService):
                         result=payload,
                     )
                 )
+                # Accumulate so the next sub-session's interim frames carry
+                # the full turn so far (used for client UI display only).
+                self._turn_committed_text += final_text
             await self._close_utterance()
 
     def _apply_recognition_result(self, recognition: dict[str, Any]) -> str: