TensorFlow lite Whisper model get worse inference result. about tensorflow HOT 2 CLOSED

class ModelClass(tf.Module):
  def __init__(self, whisper_model, *args, **kwargs):
    super().__init__(*args, **kwargs)

    self.model = whisper_model
  
  @tf.function(
      input_signature=(
        tf.TensorSpec((1, None, None), dtype=tf.float32, name="input_features"),
      )
  )
  def call(self, intput_features):
    return {"output_ids": self.model.generate(input_features=intput_features)}
  
  def get_config(self):
    return dict()
  
wrapper = ModelClass(tf_model)

# Convert mode lo tflite with dynamic quantization
converter = tf.lite.TFLiteConverter.from_keras_model(wrapper)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_ops = [
    tf.lite.OpsSet.TFLITE_BUILTINS,
    tf.lite.OpsSet.SELECT_TF_OPS,
]
tflite_quant_model = converter.convert()
open(tf_model_path , "wb").write(tflite_quant_model)