Use ndimage.median_filter instead of signal.medfilter (#812)

For a 30s long audio file which didn't have any silence, ndimage.median_filter took 7s where signa.medfilter took 30s.

Co-authored-by: Umar Farooqi <umar@paystash.com>
Co-authored-by: Jong Wook Kim <jongwook@nyu.edu>

notebooks/Multilingual_ASR.ipynb

@@ -874,7 +874,7 @@
         "from IPython.display import display, HTML\n",
         "from whisper.tokenizer import get_tokenizer\n",
         "from dtw import dtw\n",
-        "from scipy.signal import medfilt\n",
+        "from scipy.ndimage import median_filter\n",
         "\n",
         "%matplotlib inline\n",
         "%config InlineBackend.figure_format = \"retina\""
@@ -3610,7 +3610,7 @@
         "\n",
         "    weights = torch.cat(QKs)  # layers * heads * tokens * frames    \n",
         "    weights = weights[:, :, :, : duration // AUDIO_SAMPLES_PER_TOKEN].cpu()\n",
-        "    weights = medfilt(weights, (1, 1, 1, medfilt_width))\n",
+        "    weights = median_filter(weights, (1, 1, 1, medfilt_width))\n",
         "    weights = torch.tensor(weights * qk_scale).softmax(dim=-1)\n",
         "    \n",
         "    w = weights / weights.norm(dim=-2, keepdim=True)\n",