fix doc of TextDecoder (#1526)

Signed-off-by: haoshengqiang <haoshengqiang@xiaohongshu.com>
Co-authored-by: haoshengqiang <haoshengqiang@xiaohongshu.com>

whisper/model.py

@@ -197,7 +197,7 @@ class TextDecoder(nn.Module):
         """
         x : torch.LongTensor, shape = (batch_size, <= n_ctx)
             the text tokens
-        xa : torch.Tensor, shape = (batch_size, n_mels, n_audio_ctx)
+        xa : torch.Tensor, shape = (batch_size, n_audio_ctx, n_audio_state)
             the encoded audio features to be attended on
         """
         offset = next(iter(kv_cache.values())).shape[1] if kv_cache else 0