From 3b714e297ae4c1e31fb7515e45798b70a29f87fc Mon Sep 17 00:00:00 2001
From: prosolis <5590409+prosolis@users.noreply.github.com>
Date: Mon, 27 Jul 2026 13:25:28 -0700
Subject: [PATCH] Fetch the Portuguese voice Piper's own downloader can't
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

piper.download_voices pastes the voice name straight into the HTTP request
line, and http.client encodes that as ASCII — so it dies with
UnicodeEncodeError on the ã in pt_PT-tugão-medium before a byte leaves the
container. Every pt_BR voice downloads fine. The failure lands precisely on
the one voice the pt-PT pair needs, and it is the *only* European voice in
the catalogue.

The entrypoint now falls back to fetching the model and its config itself
with the path percent-encoded, which is all the downloader was missing.
Same host, same files, same destination names, so the cached-voice check
and the server invocation are unchanged.
---
 deploy/piper/entrypoint.sh | 37 ++++++++++++++++++++++++++++++++++++-
 1 file changed, 36 insertions(+), 1 deletion(-)

diff --git a/deploy/piper/entrypoint.sh b/deploy/piper/entrypoint.sh
index dc84ce4..d51caf7 100755
--- a/deploy/piper/entrypoint.sh
+++ b/deploy/piper/entrypoint.sh
@@ -10,7 +10,42 @@ port="${PIPER_PORT:-5000}"
 
 if [ ! -f "${data_dir}/${voice}.onnx" ]; then
   echo ">> downloading voice ${voice} into ${data_dir}"
-  python -m piper.download_voices "${voice}" --data-dir "${data_dir}"
+  # piper.download_voices cannot fetch a voice whose name isn't ASCII, and the
+  # only European Portuguese voice in the catalogue is pt_PT-tugão-medium:
+  # the downloader pastes the name straight into the request line, and
+  # http.client encodes that as ASCII, so it dies with UnicodeEncodeError on
+  # the ã before a byte leaves the container. Every pt_BR voice downloads
+  # fine — the failure lands precisely on the voice the pt-PT pair needs.
+  #
+  # So: try the supported path, and fall back to fetching the two files
+  # ourselves with the URL percent-encoded, which is all the downloader was
+  # missing. Same host, same files, same destination names.
+  python -m piper.download_voices "${voice}" --data-dir "${data_dir}" || {
+    echo ">> download_voices failed for ${voice}; fetching directly (non-ASCII voice name)"
+    python - "${voice}" "${data_dir}" <<'PY'
+import json, sys, urllib.parse, urllib.request
+
+voice, data_dir = sys.argv[1], sys.argv[2]
+BASE = "https://huggingface.co/rhasspy/piper-voices/resolve/main/"
+
+catalogue = json.load(urllib.request.urlopen(BASE + "voices.json", timeout=120))
+entry = catalogue.get(voice)
+if entry is None:
+    sys.exit(f"no voice named {voice!r} in the catalogue")
+
+# The catalogue keys the files by repo path; only the model and its config are
+# needed to serve (MODEL_CARD is licence text).
+for path in entry["files"]:
+    if not path.endswith((".onnx", ".onnx.json")):
+        continue
+    url = BASE + urllib.parse.quote(path)
+    dest = f"{data_dir}/{path.rsplit('/', 1)[-1]}"
+    print(f">> {url} -> {dest}", flush=True)
+    with urllib.request.urlopen(url, timeout=600) as r, open(dest, "wb") as out:
+        while chunk := r.read(1 << 20):
+            out.write(chunk)
+PY
+  }
 fi
 
 echo ">> serving ${voice} on :${port}"