diff --git a/rtcllm/a2a.py b/rtcllm/a2a.py
index 6bf836c..71169ec 100644
--- a/rtcllm/a2a.py
+++ b/rtcllm/a2a.py
@@ -1,18 +1,19 @@
 import io
-from mini_omni.inference import OmniInference
 from av import AudioFrame
 
 from aiortc import VideoStreamTrack, AudioStreamTrack
 
 class LLMAudioStreamTrack(AudioStreamTrack):
-	def __init__(self, ckpt_dir):
+	def __init__(self, omni_infer):
 		super().__init__()
 		self.ckpt_dir = ckpt_dir
-		self.oi = OmniInference(ckpt_dir=ckpt_dir)
+		# self.oi = OmniInference(ckpt_dir=ckpt_dir)
+		self.oi = omni_infer
 		self.audio_iters = []
 		self.cur_iters = None
 	
 	async def recv(self):
+		print(f'LLMAudioStreamTrack():recv() called ....')
 		b = self.get_audio_bytes()
 		if b is None:
 			return b
@@ -27,17 +28,18 @@ class LLMAudioStreamTrack(AudioStreamTrack):
 		return True
 
 	def get_audio_bytes(self):
-			if self.cur_iters is None:
-				if not set_cur_audio_iter():
-					return None
-			try:
-				b = next(self.cur_iters)
-				return b
-			except StopIteration:
-				self.cur_iters = None
-				return self.get_audio_bytes()
+		if self.cur_iters is None:
+			if not set_cur_audio_iter():
+				return None
+		try:
+			b = next(self.cur_iters)
+			return b
+		except StopIteration:
+			self.cur_iters = None
+			return self.get_audio_bytes()
 
 	def _feed(self, audio_file):
 		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 a96227b..1f96313 100644
--- a/rtcllm/rtc.py
+++ b/rtcllm/rtc.py
@@ -50,8 +50,8 @@ class RTCLLM:
 	def __init__(self, ws_url, iceServers):
 		# self.stt_model = WhisperModel('large-v3', device="cuda", compute_type="float16")
 		self.ws_url = ws_url
-		self.llmtrack = LLMAudioStreamTrack('/d/models/mini-omni')
-		self.feed = awaitify(self.llmtrack._feed)
+		self.omni_infer = 
+		self.omni_infer = OmniInference(ckpt_dir='/d/models/mini-omni')
 		self.iceServers = iceServers
 		self.peers = DictObject()
 		self.vid = 0
@@ -108,15 +108,19 @@ class RTCLLM:
 	
 	async def vad_voiceend(self, peer, audio):
 		if audio is not None:
-			ret = await self.feed(audio)
+			feed = awaitify(peer.llmtrack._feed)
+			ret = await feed(audio)
 			print(f'self.feed("{audio}") return {ret}')
 			os.remove(audio)
 
 	async def auto_accept_call(self, data):
 		opts = DictObject(iceServers=self.iceServers)
 		pc = RTCPeerConnection(opts)
+		llmtrack = LLMAudioStreamTrack('/d/models/mini-omni')
+		feed = awaitify(self.llmtrack._feed)
 		self.peers[data['from'].id] = DictObject(**{
 			'info':data['from'],
+			'llmtrack':llmtrack,
 			'pc':pc
 		})
 		pc.addTrack(self.llmtrack)