diff --git a/src/features/play-agent-audio/ui/agent-audio-player.tsx b/src/features/play-agent-audio/ui/agent-audio-player.tsx
index d422655..4bf43c1 100644
--- a/src/features/play-agent-audio/ui/agent-audio-player.tsx
+++ b/src/features/play-agent-audio/ui/agent-audio-player.tsx
@@ -1,5 +1,4 @@
 import type { Room } from "livekit-client";
-import { Button } from "@/shared/ui";
 import { usePlayAgentAudio } from "../model/use-play-agent-audio";
 
 export function AgentAudioPlayer({ room }: { room: Room | null }) {
@@ -9,9 +8,19 @@ export function AgentAudioPlayer({ room }: { room: Room | null }) {
 			{/* biome-ignore lint/a11y/useMediaCaption: synthesized speech has no source track to caption */}
 			<audio autoPlay playsInline ref={audioRef} />
 			{needsPlayTap && (
-				<Button onClick={retryPlay} size="sm" variant="outline">
-					소리를 재생하려면 눌러주세요
-				</Button>
+				// Browsers (iOS Safari in particular) block audio.play() unless it's called
+				// directly from a user gesture, so the initial autoplay attempt in
+				// use-play-agent-audio.ts fails silently -- without a hard-to-miss prompt, the
+				// call just looks completely silent with no indication why. Full-width and
+				// high-contrast on purpose: this is the only control on the HEARING screen, and
+				// missing it means the whole call is silent for that person.
+				<button
+					type="button"
+					onClick={retryPlay}
+					className="w-full max-w-sm animate-pulse rounded-2xl bg-destructive px-5 py-4 text-center text-base font-bold text-white shadow-lg focus-visible:outline-4 focus-visible:outline-offset-2 focus-visible:outline-ring"
+				>
+					소리가 재생되지 않고 있어요 — 여기를 눌러 소리 켜기
+				</button>
 			)}
 		</>
 	);
diff --git a/src/features/sign-capture/model/use-sign-capture.ts b/src/features/sign-capture/model/use-sign-capture.ts
index fdf3fde..d57df82 100644
--- a/src/features/sign-capture/model/use-sign-capture.ts
+++ b/src/features/sign-capture/model/use-sign-capture.ts
@@ -478,7 +478,12 @@ export function useSignCapture(
 				}
 				onSentence?.(sentence);
 			})
-			.catch(() => {
+			.catch((error) => {
+				// Falling back to the raw joined words is intentional (never drop the signer's
+				// message), but doing it silently makes a failed compose call indistinguishable
+				// from the LLM genuinely returning the words unchanged -- log it so a network/API
+				// failure is diagnosable instead of just looking like "GPT isn't running".
+				console.error("[sign-capture] compose failed, sending raw words instead:", error);
 				const fallback = words.join(" ");
 				setComposedSentences((prev) => [...prev, fallback].slice(-MAX_DISPLAYED_SENTENCES));
 				const turn: ConversationTurn = { speaker: "DEAF", text: fallback };