diff --git a/sleap/io/format/hdf5.py b/sleap/io/format/hdf5.py
index 353f88e3a..55a30d74f 100644
--- a/sleap/io/format/hdf5.py
+++ b/sleap/io/format/hdf5.py
@@ -81,7 +81,10 @@ def read_headers(
 
         # Extract the Labels JSON metadata and create Labels object with just this
         # metadata.
-        dicts = json_loads(f.require_group("metadata").attrs["json"].tobytes().decode())
+        json = f.require_group("metadata").attrs["json"]
+        if not isinstance(json, str):
+            json = json.tobytes().decode()
+        dicts = json_loads(json)
 
         # These items are stored in separate lists because the metadata group got to be
         # too big.
@@ -151,6 +154,45 @@ def read(
             points_dset[:]["x"] -= 0.5
             points_dset[:]["y"] -= 0.5
 
+        def cast_as_compound(arr, dtype):
+            out = np.empty(shape=(len(arr),), dtype=dtype)
+            if out.size == 0:
+                return out
+            for i, (name, _) in enumerate(dtype):
+                out[name] = arr[:, i]
+            return out
+
+        # cast points, instances, and frames into complex dtype if not already
+        dtype_points = [("x", "<f8"), ("y", "<f8"), ("visible", "?"), ("complete", "?")]
+        if points_dset.dtype.kind != "V":
+            points_dset = cast_as_compound(points_dset, dtype_points)
+        if pred_points_dset.dtype.kind != "V":
+            pred_points_dset = cast_as_compound(pred_points_dset, dtype_points)
+
+        dtype_instances = [
+            ("instance_id", "<i8"),
+            ("instance_type", "u1"),
+            ("frame_id", "<u8"),
+            ("skeleton", "<u4"),
+            ("track", "<i4"),
+            ("from_predicted", "<i8"),
+            ("score", "<f4"),
+            ("point_id_start", "<u8"),
+            ("point_id_end", "<u8"),
+        ]
+        if instances_dset.dtype.kind != "V":
+            instances_dset = cast_as_compound(instances_dset, dtype_instances)
+
+        dtype_frames = [
+            ("frame_id", "<u8"),
+            ("video", "<u4"),
+            ("frame_idx", "<u8"),
+            ("instance_id_start", "<u8"),
+            ("instance_id_end", "<u8"),
+        ]
+        if frames_dset.dtype.kind != "V":
+            frames_dset = cast_as_compound(frames_dset, dtype_frames)
+
         # Rather than instantiate a bunch of Point\PredictedPoint objects, we will use
         # inplace numpy recarrays. This will save a lot of time and memory when reading
         # things in.
@@ -283,9 +325,10 @@ def write(
             if append and "json" in meta_group.attrs:
 
                 # Otherwise, we need to read the JSON and append to the lists
-                old_labels = labels_json.LabelsJsonAdaptor.from_json_data(
-                    meta_group.attrs["json"].tobytes().decode()
-                )
+                json = meta_group.attrs["json"]
+                if not isinstance(json, str):
+                    json = json.tobytes().decode()
+                old_labels = labels_json.LabelsJsonAdaptor.from_json_data(json)
 
                 # A function to join to list but only include new non-dupe entries
                 # from the right hand list.