diff --git a/testing/test_songs.py b/testing/test_songs.py
new file mode 100644
index 0000000..702c8ad
--- /dev/null
+++ b/testing/test_songs.py
@@ -0,0 +1,100 @@
+import os
+import shutil
+import zipfile
+import re
+
+import pytest
+
+from tja2fumen import main as convert
+from tja2fumen.parsers import readFumen
+from tja2fumen.constants import COURSE_IDS, NORMALIZE_COURSE
+
+
+def normalize_type(note_type):
+    return re.sub(r'[0-9]', '', note_type)
+
+
+def assert_song_property(obj1, obj2, prop, measure=None, branch=None, note=None, func=None, abs=None):
+    # NB: TJA parser/converter uses 0-based indexing, but TJA files use 1-based indexing.
+    #     So, we increment 1 in the error message to more easily identify problematic lines in TJA files.
+    msg_failure = f"'{prop}' mismatch"
+    msg_failure += f": measure '{measure+1}'" if measure is not None else ""
+    msg_failure += f", branch '{branch}'" if branch is not None else ""
+    msg_failure += f", note '{note+1}'" if note is not None else ""
+    if func:
+        assert func(obj1[prop]) == func(obj2[prop]), msg_failure
+    elif abs:
+        assert obj1[prop] == pytest.approx(obj2[prop], abs=abs), msg_failure
+    else:
+        assert obj1[prop] == obj2[prop], msg_failure
+
+
+@pytest.mark.parametrize('id_song', ['song1', 'song2'])
+def test_converted_tja_vs_cached_fumen(id_song, tmp_path):
+    # Define the testing directory
+    path_test = os.path.dirname(os.path.realpath(__file__))
+
+    # Define the working directory
+    path_temp = os.path.join(tmp_path, id_song)
+    os.mkdir(path_temp)
+
+    # Copy input TJA to working directory
+    path_tja = os.path.join(path_test, "data", f"{id_song}.tja")
+    path_tja_tmp = os.path.join(path_temp, f"{id_song}.tja")
+    shutil.copy(path_tja, path_tja_tmp)
+
+    # Convert TJA file to fumen files
+    _, _, paths_out = convert(argv=[path_tja_tmp])
+
+    # Extract cached fumen files to working directory
+    path_binzip = os.path.join(path_test, "data", f"{id_song}.zip")
+    path_bin = os.path.join(path_temp, "ca_bins")
+    with zipfile.ZipFile(path_binzip, 'r') as zip_ref:
+        zip_ref.extractall(path_bin)
+
+    # Compare cached fumen with generated fumen
+    for path_out in paths_out:
+        # Difficulty introspection to help with debugging
+        i_difficult_id = os.path.basename(path_out).split(".")[0].split("_")[1]
+        i_difficulty = NORMALIZE_COURSE[{v: k for k, v in COURSE_IDS.items()}[i_difficult_id]]  # noqa
+        # 0. Read fumen data (converted vs. cached)
+        co_song = readFumen(path_out)
+        ca_song = readFumen(os.path.join(path_bin, os.path.basename(path_out)))
+        # 1. Check song headers
+        assert_song_property(co_song, ca_song, 'header', func=len)
+        assert_song_property(co_song, ca_song, 'headerUnknown', func=len)
+        # 2. Check song metadata
+        assert_song_property(co_song, ca_song, 'order')
+        assert_song_property(co_song, ca_song, 'branches')
+        assert_song_property(co_song, ca_song, 'scoreInit')
+        assert_song_property(co_song, ca_song, 'scoreDiff')
+        assert_song_property(co_song, ca_song, 'length')
+        assert_song_property(co_song, ca_song, 'measures', func=len)
+        # 3. Check measure data
+        for i_measure, (co_measure, ca_measure) in enumerate(zip(co_song['measures'], ca_song['measures'])):
+            # 3a. Check measure metadata
+            assert_song_property(co_measure, ca_measure, 'bpm', i_measure, abs=0.01)
+            assert_song_property(co_measure, ca_measure, 'fumenOffset', i_measure, abs=0.5)
+            assert_song_property(co_measure, ca_measure, 'gogo', i_measure)
+            assert_song_property(co_measure, ca_measure, 'barline', i_measure)
+            # 3b. Check measure notes
+            for i_branch in ['normal', 'advanced', 'master']:
+                co_branch = co_measure[i_branch]
+                ca_branch = ca_measure[i_branch]
+                assert_song_property(co_branch, ca_branch, 'length', i_measure, i_branch)
+                # NB: We check for branching before checking speed as fumens store speed changes even for empty branches
+                if co_branch['length'] == 0:
+                    continue
+                assert_song_property(co_branch, ca_branch, 'speed', i_measure, i_branch)
+                for i_note in range(co_branch['length']):
+                    co_note = co_branch[i_note]
+                    ca_note = ca_branch[i_note]
+                    assert_song_property(co_note, ca_note, 'type', i_measure, i_branch, i_note, func=normalize_type)
+                    assert_song_property(co_note, ca_note, 'pos', i_measure, i_branch, i_note, abs=0.1)
+                    # NB: Drumroll duration doesn't always end exactly on a beat. So, use a larger tolerance.
+                    assert_song_property(co_note, ca_note, 'duration', i_measure, i_branch, i_note, abs=10.0)
+                    if ca_note['type'] not in ["Balloon", "Kusudama"]:
+                        assert_song_property(co_note, ca_note, 'scoreInit', i_measure, i_branch, i_note)
+                        assert_song_property(co_note, ca_note, 'scoreDiff', i_measure, i_branch, i_note)
+                    # assert_song_property(co_note, ca_note, 'item', i_measure, i_branch, i_note)
+