diff --git a/nginx/snippets/proxy-pass.conf b/nginx/snippets/proxy-pass.conf
index 06ffe1a..b84b86a 100644
--- a/nginx/snippets/proxy-pass.conf
+++ b/nginx/snippets/proxy-pass.conf
@@ -4,11 +4,8 @@
 set $upstream_app app:8000;
 proxy_pass http://$upstream_app;
 
-# No backend synthesis timeout exists today (ffmpeg.py only bounds the
-# encode step at 30s; TTS inference itself is unbounded) — pick a ceiling
-# safely above realistic worst-case rather than nginx's low defaults cutting
-# off legitimate slow requests.
-proxy_read_timeout 120s;
+# proxy_read_timeout is set per location (nginx rejects a directive repeated in
+# one block, so it can't live here and be overridden after the include).
 proxy_send_timeout 120s;
 proxy_connect_timeout 5s;
 
diff --git a/nginx/templates/20-speech-server.conf.template b/nginx/templates/20-speech-server.conf.template
index 7577de2..d8aa73f 100644
--- a/nginx/templates/20-speech-server.conf.template
+++ b/nginx/templates/20-speech-server.conf.template
@@ -74,9 +74,8 @@ server {
         include /etc/nginx/snippets/proxy-pass.conf;
 
         # TTS inference is CPU-bound and can be slow (24-layer models, long text,
-        # emulated arch) — give synthesis more than the shared 120s read timeout so
-        # a legitimately long generation isn't cut off with a 504. This override must
-        # come AFTER the include, whose 120s would otherwise win.
+        # emulated arch) — give synthesis more than the 120s default on other
+        # routes so a legitimately long generation isn't cut off with a 504.
         proxy_read_timeout 300s;
     }
 
@@ -91,5 +90,9 @@ server {
         limit_conn perip 20;
 
         include /etc/nginx/snippets/proxy-pass.conf;
+
+        # No backend timeout exists today (TTS inference is unbounded) — pick a
+        # ceiling safely above realistic worst-case rather than nginx's low default.
+        proxy_read_timeout 120s;
     }
 }