diff --git a/rtcllm/a2a.py b/rtcllm/a2a.py
index d44f0ec..3d7957d 100644
--- a/rtcllm/a2a.py
+++ b/rtcllm/a2a.py
@@ -1,3 +1,4 @@
+import os
 from traceback import print_exc
 import io
 import asyncio
@@ -11,13 +12,12 @@ class LLMAudioStreamTrack(AudioStreamTrack):
 		self.oi = omni_infer
 		self.audio_iters = []
 		self.cur_iters = None
+		self.tmp_files = []
 	
 	async def recv(self):
-		print(f'LLMAudioStreamTrack():recv() called ....')
 		try:
 			b = self.get_audio_bytes()
 			if b is None:
-				print('LLMAudioStreamTrack return None ...')
 				return await super().recv()
 			frame = AudioFrame.from_ndarray(io.BytesIO(b), format='s16', layout='mono')
 			print('LLMAudioStreamTrack return frame ...')
@@ -43,9 +43,17 @@ class LLMAudioStreamTrack(AudioStreamTrack):
 			return b
 		except StopIteration:
 			self.cur_iters = None
+			if len(self.tmp_files) > 0:
+				tf = self.tmp_files[0]
+				self.tmp_files.remove(tf)
+				os.remove(tf)
 			return self.get_audio_bytes()
 
 	def _feed(self, audio_file):
+		self.tmp_files.append(audio_file)
+		if audio_file is None:
+			print(f'*****{self.__class__.__name__}._feed(),{audio_file=}')
+			return
 		x = self.oi.run_AT_batch_stream(audio_file)
 		self.audio_iters.append(x)
 		return x
diff --git a/rtcllm/rtc.py b/rtcllm/rtc.py
index 4e1035a..685c3ee 100644
--- a/rtcllm/rtc.py
+++ b/rtcllm/rtc.py
@@ -113,7 +113,7 @@ class RTCLLM:
 			feed = awaitify(peer.llmtrack._feed)
 			ret = await feed(audio)
 			print(f'self.feed("{audio}") return {ret}')
-			os.remove(audio)
+			# os.remove(audio)
 
 	async def auto_accept_call(self, data):
 		opts = DictObject(iceServers=self.iceServers)