try

app.py

@@ -35,15 +35,17 @@ def transcribe(audio: tuple[int, np.ndarray], conversation: list[dict], gradio_c
 
     audio_sr = librosa.resample(audio[1].astype(np.float32) / 32768.0,
                                 orig_sr=original_sr, target_sr=target_sr)
+    inputs = pipe.preprocess({"audio": audio_sr, "turns": conversation, "sampling_rate": target_sr})
+    print("preprocess", inputs)
    
     output = pipe({"audio": audio_sr, "turns": conversation, "sampling_rate": target_sr},
                   max_new_tokens=512)
     print("output", output)
 
-    conversation.append({"role": "user", "content": output["transcription"]})
-    conversation.append({"role": "assistant", "content": output["reply"]})
-    gradio_convo.append({"role": "user", "content": output["transcription"]})
-    gradio_convo.append({"role": "assistant", "content": output["reply"]})
+    conversation.append({"role": "user", "content": "hi"})
+    conversation.append({"role": "assistant", "content": output})
+    gradio_convo.append({"role": "user", "content": "hi"})
+    gradio_convo.append({"role": "assistant", "content": output})
 
     yield AdditionalOutputs(conversation, gradio_convo)