diff --git a/src/transformers/models/whisper/feature_extraction_whisper.py b/src/transformers/models/whisper/feature_extraction_whisper.py
index 4151a3824dfd..8adbf8fa20e7 100644
--- a/src/transformers/models/whisper/feature_extraction_whisper.py
+++ b/src/transformers/models/whisper/feature_extraction_whisper.py
@@ -147,7 +147,11 @@ def _torch_extract_fbank_features(self, waveform: np.ndarray, device: str = "cpu
             waveform += self.dither * torch.randn(waveform.shape, dtype=waveform.dtype, device=waveform.device)
 
         stft = torch.stft(waveform, self.n_fft, self.hop_length, window=window, return_complex=True)
-        magnitudes = stft[..., :-1].abs() ** 2
+        # `stft[..., :-1]` is a non-contiguous view; on some CPU backends the
+        # downstream `mel_filters.T @ magnitudes` matmul hits a slow strided
+        # path (observed ~8x slower). Forcing contiguity here is cheap and
+        # keeps the matmul on the fast path.
+        magnitudes = (stft[..., :-1].abs() ** 2).contiguous()
 
         mel_filters = torch.from_numpy(self.mel_filters).to(device, torch.float32)
         mel_spec = mel_filters.T @ magnitudes
diff --git a/tests/models/whisper/test_feature_extraction_whisper.py b/tests/models/whisper/test_feature_extraction_whisper.py
index 65f150cc1295..f7a9d590619e 100644
--- a/tests/models/whisper/test_feature_extraction_whisper.py
+++ b/tests/models/whisper/test_feature_extraction_whisper.py
@@ -268,6 +268,45 @@ def test_double_precision_pad(self):
             pt_processed = feature_extractor.pad([{"input_features": inputs}], return_tensors="pt")
             self.assertTrue(pt_processed.input_features.dtype == torch.float32)
 
+    @require_torch
+    def test_torch_extract_fbank_features_contiguous_magnitudes(self):
+        """Guards the mel-magnitude contiguity fix.
+
+        `stft[..., :-1]` is a non-contiguous view, and squaring it keeps that
+        layout. On some CPU backends the downstream `mel_filters.T @ magnitudes`
+        matmul then falls onto a slow strided-GEMM path (observed ~8x slower on
+        a ROCm PyTorch build, ~43 ms vs ~3 ms for a 30 s window). Forcing
+        `magnitudes` contiguous restores the fast path with no change in output.
+
+        Correctness (and the non-contiguity of the raw view) is asserted so this
+        stays stable in CI. See https://github.com/huggingface/transformers/pull/47351
+        for the benchmark script and the measured (backend-dependent) speedups.
+        """
+        torch.manual_seed(0)
+        feature_extractor = WhisperFeatureExtractor()
+        n_fft = feature_extractor.n_fft
+        hop_length = feature_extractor.hop_length
+        # Mirror `_torch_extract_fbank_features`: the window is built from n_fft.
+        window = torch.hann_window(n_fft)
+        # One full Whisper window of audio (chunk_length * sampling_rate samples).
+        waveform = torch.randn(feature_extractor.n_samples, dtype=torch.float32)
+
+        stft = torch.stft(waveform, n_fft, hop_length, window=window, return_complex=True)
+        magnitudes_non_contiguous = stft[..., :-1].abs() ** 2
+        magnitudes_contiguous = magnitudes_non_contiguous.contiguous()
+
+        # Root cause of the slow path: the sliced view is non-contiguous.
+        self.assertFalse(magnitudes_non_contiguous.is_contiguous())
+        self.assertTrue(magnitudes_contiguous.is_contiguous())
+
+        mel_filters = torch.from_numpy(feature_extractor.mel_filters).to(torch.float32)
+
+        # The fix must not change the result, only the memory layout.
+        torch.testing.assert_close(
+            mel_filters.T @ magnitudes_non_contiguous,
+            mel_filters.T @ magnitudes_contiguous,
+        )
+
     def _load_datasamples(self, num_samples):
         ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
         # automatic decoding with librispeech