Reduce precision difference for whipser model (#4116)

2 days ago · af2bdf120f
parent 1e3e186c18
commit af2bdf120f
1 changed files with 3 additions and 1 deletions
--- 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()