diff --git a/piper/CHANGELOG.md b/piper/CHANGELOG.md
index a44e8a729..0b30ca47c 100644
--- a/piper/CHANGELOG.md
+++ b/piper/CHANGELOG.md
@@ -1,5 +1,14 @@
 # Changelog
 
+## 2.5.3
+
+- Voice models downloaded from Piper's catalog are left out of backups again.
+  They are moved to `/data/catalog_voices` at startup and excluded from there,
+  which keeps a catalog voice (20-110 MB each) out of every backup while
+  uploaded voices, which exist nowhere else, stay in. A model is only moved
+  when its md5 matches the catalog's, so a custom voice that reuses a catalog
+  name is still backed up
+
 ## 2.5.2
 
 - Upgrade to `wyoming-piper` 2.5.2
diff --git a/piper/DOCS.md b/piper/DOCS.md
index 8d3a3ab7a..ae9e11363 100644
--- a/piper/DOCS.md
+++ b/piper/DOCS.md
@@ -135,9 +135,18 @@ interface, because that directory is mounted read-only. Delete those files
 directly instead.
 
 Because voices can be added and removed here, the set you end up with is your
-own and is included in the app's backups — including uploads, which exist
-nowhere else. Only the large re-downloadable caches are left out: the OmniVoice
-model weights and the downloads for the Japanese and Thai options.
+own, and an uploaded voice exists nowhere else, so uploads are included in the
+app's backups. What can be fetched again is not: a voice from Piper's catalog
+is moved to `/data/catalog_voices` when the app starts and left out of backups
+from then on, as are the OmniVoice model weights and the downloads for the
+Japanese and Thai options. A restore downloads the catalog voices it needs
+again, on demand.
+
+The move only happens when the file is byte for byte the catalog's, so a voice
+of your own that borrows a catalog name — a fine-tune exported as
+`en_US-lessac-medium.onnx`, say — keeps its place in backups. A voice already
+downloaded while the app is running is moved at the next restart, so one backup
+can still include it.
 
 ## Custom Voices
 
diff --git a/piper/config.yaml b/piper/config.yaml
index 0da7732cd..5e228107d 100644
--- a/piper/config.yaml
+++ b/piper/config.yaml
@@ -1,5 +1,5 @@
 ---
-version: 2.5.2
+version: 2.5.3
 slug: piper
 name: Piper
 description: Text-to-speech with Piper
@@ -10,11 +10,11 @@ arch:
 init: false
 discovery:
   - wyoming
-# Voice models ("*.onnx") used to be excluded, back when /data held nothing but
-# re-downloadable copies of catalog voices. The web interface now uploads and
-# deletes voices there, so which models are present is the user's own choice and
-# has to survive a backup -- an uploaded voice exists nowhere else. Only the
-# genuinely disposable caches below are skipped.
+# Voice models ("*.onnx") used to be excluded wholesale, back when /data held
+# nothing but re-downloadable copies of catalog voices. The web interface now
+# uploads and deletes voices there, so which models are present is the user's
+# own choice and has to survive a backup -- an uploaded voice exists nowhere
+# else. Only what can be fetched again is skipped.
 backup_exclude:
   # OmniVoice weights, 632 MB. The 1.5 MB graph beside them is kept, but the
   # backend needs both files, so a restore without the weights just downloads
@@ -24,6 +24,11 @@ backup_exclude:
   - "*/hub"
   # Wheel cache for the optional packs, rebuilt on demand.
   - "*/uv-cache"
+  # Voice models that came from the catalog, moved here at startup by
+  # sort_catalog_voices.py precisely so they can be excluded. A model is only
+  # put here when its md5 matches the catalog's, which makes it a copy of a
+  # file that is still published; uploads stay in /data and stay in backups.
+  - "*/catalog_voices"
 map:
   - share
 ingress: true
diff --git a/piper/rootfs/etc/s6-overlay/s6-rc.d/piper/run b/piper/rootfs/etc/s6-overlay/s6-rc.d/piper/run
index 75f910114..708044fd3 100755
--- a/piper/rootfs/etc/s6-overlay/s6-rc.d/piper/run
+++ b/piper/rootfs/etc/s6-overlay/s6-rc.d/piper/run
@@ -11,6 +11,13 @@
 # <language>/<name>/ directory instead.
 readonly omnivoice_voices_dir='/data/omnivoice_voices'
 
+# Voice models that came from the catalog are kept apart from uploaded ones, so
+# config.yaml can leave the re-downloadable half out of backups. wyoming-piper
+# writes both kinds to --download-dir, which stays /data: an upload that is not
+# backed up yet is a file that can be lost, while a catalog voice that is still
+# in /data only makes the next backup bigger until it is sorted below.
+readonly catalog_voices_dir='/data/catalog_voices'
+
 # Ingress proxies to the container's address rather than to loopback, so the web
 # UI has to bind every interface -- which also puts it in reach of every other
 # app on the same Docker network. It has no authentication of its own, so the
@@ -33,6 +40,15 @@ fi
 
 mkdir -p "${omnivoice_voices_dir}"
 
+# Before Piper starts, so a voice is never moved out from under it. Never
+# fatal: an unsorted voice costs backup space and nothing else.
+if ! sorted_voices="$(python3 /usr/local/bin/sort_catalog_voices.py \
+    /data "${catalog_voices_dir}" 2>&1)"; then
+    bashio::log.warning "Could not sort catalog voices: ${sorted_voices}"
+elif bashio::var.has_value "${sorted_voices}"; then
+    bashio::log.info "${sorted_voices}"
+fi
+
 backend="$(bashio::config 'backend')"
 if [[ "${backend}" == 'omnivoice' ]]; then
     if ! have_module 'omnivoice'; then
@@ -94,6 +110,7 @@ exec python3 -m wyoming_piper \
     --zeroconf \
     --sentence-silence "$(bashio::config 'sentence_silence')" \
     --data-dir /data \
+    --data-dir "${catalog_voices_dir}" \
     --data-dir /share/piper \
     --download-dir /data \
     --omnivoice-ref-dir "${omnivoice_voices_dir}" \
diff --git a/piper/rootfs/usr/local/bin/sort_catalog_voices.py b/piper/rootfs/usr/local/bin/sort_catalog_voices.py
new file mode 100644
index 000000000..970e97364
--- /dev/null
+++ b/piper/rootfs/usr/local/bin/sort_catalog_voices.py
@@ -0,0 +1,96 @@
+"""Move re-downloadable voice models out of the backed-up part of /data.
+
+Catalog voices and uploaded ones sit side by side: wyoming-piper downloads
+into --download-dir, and the web interface uploads there too. Only the uploads
+are irreplaceable, while a single catalog voice is 20-110 MB of every backup
+the add-on is part of, for something a restore fetches again on demand.
+
+So the catalog copies are moved into a subdirectory that config.yaml leaves
+out of backups, and that subdirectory is handed back to wyoming-piper as
+another --data-dir. Nothing is re-downloaded now; after a restore the
+subdirectory is simply absent, which is the state of a fresh install.
+
+A file is moved only when its md5 matches what the catalog lists for that
+name, so a voice that merely borrows a catalog name -- a fine-tune exported as
+en_US-lessac-medium.onnx, say -- stays in /data and stays in backups. Names
+are checked before hashing, so the common case of a custom voice under a
+custom name reads no bytes at all.
+
+This runs at startup because the add-on does not control where wyoming-piper
+writes: a voice downloaded while the add-on is running stays in /data until
+the next start and is moved then.
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List, Set
+
+from wyoming_piper.download import get_voices
+from wyoming_piper.file_hash import get_file_hash
+
+
+def catalog_digests(download_dir: Path) -> Dict[str, Set[str]]:
+    """Map file name to the md5 digests the catalog publishes for that name.
+
+    A set rather than one digest: the same file is listed under a voice and
+    under each of its aliases, and nothing guarantees those agree.
+    """
+    digests: Dict[str, Set[str]] = {}
+
+    # update_voices is left off: this must not depend on the network, and the
+    # Piper service downloads the current voices.json moments later anyway.
+    for voice in get_voices(download_dir).values():
+        for file_path, file_info in voice.get("files", {}).items():
+            digest = file_info.get("md5_digest")
+            if digest:
+                digests.setdefault(Path(file_path).name, set()).add(digest)
+
+    return digests
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("data_dir", help="Directory voices are downloaded into")
+    parser.add_argument("catalog_dir", help="Directory to move catalog voices to")
+    args = parser.parse_args()
+
+    data_dir = Path(args.data_dir)
+    catalog_dir = Path(args.catalog_dir)
+    digests = catalog_digests(data_dir)
+
+    moved: List[str] = []
+    moved_bytes = 0
+
+    for onnx_path in sorted(data_dir.glob("*.onnx")):
+        # A voice is the model and its config together. Moving one without the
+        # other would hide the voice from both directories.
+        pair = [onnx_path, onnx_path.with_name(f"{onnx_path.name}.json")]
+
+        if not all(path.is_file() and not path.is_symlink() for path in pair):
+            continue
+
+        if not all(path.name in digests for path in pair):
+            continue
+
+        if not all(get_file_hash(path) in digests[path.name] for path in pair):
+            continue
+
+        catalog_dir.mkdir(parents=True, exist_ok=True)
+        for path in pair:
+            moved_bytes += path.stat().st_size
+            # Same filesystem, so this is a rename: the voice is in exactly one
+            # of the two directories at every point, never half-copied.
+            path.replace(catalog_dir / path.name)
+
+        moved.append(onnx_path.name[: -len(".onnx")])
+
+    if moved:
+        print(
+            f"Moved {len(moved)} re-downloadable voice(s), "
+            f"{moved_bytes / (1024 * 1024):.0f} MiB, to {catalog_dir}, "
+            f"which is left out of backups: {', '.join(moved)}"
+        )
+
+
+if __name__ == "__main__":
+    main()