fix: move numpy and soundfile imports inside transcribe function and clean up audio buffer handling

app.py

@@ -23,19 +23,17 @@ def load_model():
         print(f"Model loaded on device: {model.device}")
     return model
 
-import numpy as np
-import soundfile as sf
-audio_buffer = []
 
 @spaces.GPU(duration=120)
 def transcribe(audio, state=""):
     # Load the model inside the GPU worker process
+    import numpy as np
+    import soundfile as sf
     model = load_model()
     
     if audio is None or isinstance(audio, int):
         print(f"Skipping invalid audio input: {type(audio)}")
         return state, state
-
     print(f"Received audio input of type: {type(audio)}")
     print(f"Audio shape: {audio.shape if isinstance(audio, np.ndarray) else 'N/A'}")
     # Append NumPy array to buffer
@@ -84,8 +82,6 @@ def transcribe(audio, state=""):
             print(f"Error processing audio: {e}")
             # return state, state
 
-        # Clear buffer
-        audio_buffer = []
         new_state = state + " " + transcription if state else transcription
         return new_state, new_state
     return state, state