diff --git a/lib/ads/video/encode.ts b/lib/ads/video/encode.ts
index aeede36..4aaf348 100644
--- a/lib/ads/video/encode.ts
+++ b/lib/ads/video/encode.ts
@@ -98,6 +98,16 @@ export function mp4Args(o: Mp4EncodeOptions): string[] {
   }
 
   if (o.audioPath) {
+    // `apad` extends the audio with silence indefinitely, and `-shortest` then
+    // trims the result to the video's five seconds — so the track is exactly as
+    // long as the picture.
+    //
+    // Without the pad, a narration shorter than the ad (which every narration
+    // is: a five-second read is about two and a half seconds of speech) left an
+    // audio stream that ended early. Validation requires the two durations to
+    // match within one AAC frame and rejected every narrated render, and a
+    // player handed a short track is entitled to stop at its end.
+    args.push("-af", "apad");
     args.push("-c:a", "aac", "-profile:a", "aac_low", "-ar", String(AAC_SAMPLE_RATE), "-b:a", "128k", "-ac", "2", "-shortest");
   } else {
     args.push("-an");
diff --git a/tests/ads-video-compose-validate.test.ts b/tests/ads-video-compose-validate.test.ts
index 6a834b4..800efd7 100644
--- a/tests/ads-video-compose-validate.test.ts
+++ b/tests/ads-video-compose-validate.test.ts
@@ -187,6 +187,23 @@ describe("encoder arguments carry the media contract", () => {
     expect(narrated[narrated.indexOf("-ar") + 1]).toBe("48000");
   });
 
+  it("pads a narration to the length of the picture", () => {
+    // Every narration is shorter than its ad — a five-second read is about two
+    // and a half seconds of speech. Without the pad the audio stream ended
+    // early, which failed the duration check on every narrated render and left
+    // players entitled to stop at the end of the track.
+    const a = mp4Args({
+      framePattern: "f-%04d.png",
+      audioPath: "/tmp/narration.mp3",
+      outPath: "/tmp/out.mp4",
+      profile: "master_1080p",
+      videoKbps: 4000,
+    }).join(" ");
+    expect(a).toContain("-af apad");
+    // apad alone runs forever; -shortest is what trims it back to the video.
+    expect(a).toContain("-shortest");
+  });
+
   it("stream-copies into HLS rather than re-encoding", () => {
     const hls = hlsArgs({ inPath: "/tmp/720.mp4", outDir: "/tmp/hls", name: "720p" });
     // Re-encoding would move the keyframes and break the boundaries the MP4 was