diff --git a/lidbox/data/steps.py b/lidbox/data/steps.py
index 2b246aa..33072cd 100644
--- a/lidbox/data/steps.py
+++ b/lidbox/data/steps.py
@@ -5,6 +5,7 @@
 """
 import collections
 import io
+import json
 import logging
 import os
 import shutil
@@ -391,6 +392,28 @@ def cache(ds, directory=None, batch_size=1, cache_key=None):
               .unbatch())
 
 
+def validate_cache(dataframe, path, cache_key):
+    """
+    Validate any existing cache. Validation is based on saving the keys and shape of the given
+    dataframe to JSON-format.
+    """
+    cache_file = f"{path}/{cache_key}_meta.json"
+    if os.path.exists(cache_file):
+        with open(cache_file, 'r', encoding='utf-8') as infile:
+            existing_values = json.load(infile)
+            new_keys = dataframe.columns.to_list()
+            assert existing_values["keys"] == new_keys, \
+                    f"Cache validation failed, old keys {existing_values['keys']} vs. new {new_keys}"
+            assert existing_values["shape"] == list(dataframe.shape), \
+                    f"Cache validation failed, old shape {existing_values['shape']} vs. new {dataframe.shape}"
+            logger.info("Cache validation passed.")
+    else:
+        values = {"keys": dataframe.columns.to_list(), "shape": dataframe.shape}
+        logger.info(f"Previous cache does not exist. Saving dataframe keys and shape to {cache_file} for validation.")
+        with open(cache_file, 'w', encoding='utf-8') as outfile:
+            json.dump(values, outfile, indent=2)
+
+
 def compute_rms_vad(ds, strength, vad_frame_length_ms, min_non_speech_length_ms=0):
     """
     Compute root mean square based voice activity detection.