diff --git a/paddlespeech/s2t/models/whisper/whisper.py b/paddlespeech/s2t/models/whisper/whisper.py
index 54aef0956..9ae2a7de5 100644
--- a/paddlespeech/s2t/models/whisper/whisper.py
+++ b/paddlespeech/s2t/models/whisper/whisper.py
@@ -1613,7 +1613,9 @@ def log_mel_spectrogram(audio: Union[str, np.ndarray, paddle.Tensor],
     magnitudes = stft[:, :-1].abs()**2
 
     filters = mel_filters(resource_path, n_mels)
-    mel_spec = filters @ magnitudes
+    mel_spec = paddle.to_tensor(
+        filters.numpy()
+        @ magnitudes.numpy())  # Use numpy to reduce precision difference
     mel_spec = paddle.to_tensor(mel_spec.numpy().tolist())
 
     log_spec = paddle.clip(mel_spec, min=1e-10).log10()