Fix TTS: switch to 1.7B with ref_audio, speakable text on all lines

- Use 1.7B model (0.6B had tensor mismatch with cached prompts)
- Speak endpoint uses ref_audio directly (not cached pkl) as fallback
- Cache voice clone prompts in memory on startup
- Add SpeakableText component: 🔊 icon on each p and li element
- Remove old TTSReader sequential approach
- Add global exception handler to TTS server
- Fix profile localStorage caching
- inference_mode + bf16 optimization

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

sundol-frontend/src/app/knowledge/[id]/page.tsx

@@ -6,6 +6,7 @@ import AuthGuard from "@/components/auth-guard";
 import NavBar from "@/components/nav-bar";
 import { useApi } from "@/lib/use-api";
 import ReactMarkdown from "react-markdown";
+import SpeakableText from "@/components/speakable-text";
 
 interface Category {
   ID: string;
@@ -314,10 +315,18 @@ export default function KnowledgeDetailPage() {
                       h1: ({children}) => <h1 className="text-xl font-bold mt-6 mb-3">{children}</h1>,
                       h2: ({children}) => <h2 className="text-lg font-bold mt-5 mb-2">{children}</h2>,
                       h3: ({children}) => <h3 className="text-base font-bold mt-4 mb-2">{children}</h3>,
-                      p: ({children}) => <p className="mb-3">{children}</p>,
+                      p: ({children, node}) => {
+                        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                        const txt = node?.children?.map((c: any) => c.type === 'text' ? c.value : '').join('') || '';
+                        return <p className="mb-3"><SpeakableText text={txt}>{children}</SpeakableText></p>;
+                      },
                       ul: ({children}) => <ul className="list-disc ml-5 mb-3 space-y-1">{children}</ul>,
                       ol: ({children}) => <ol className="list-decimal ml-5 mb-3 space-y-1">{children}</ol>,
-                      li: ({children}) => <li className="leading-relaxed">{children}</li>,
+                      li: ({children, node}) => {
+                        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                        const txt = node?.children?.map((c: any) => c.type === 'text' ? c.value : '').join('') || '';
+                        return <li className="leading-relaxed"><SpeakableText text={txt}>{children}</SpeakableText></li>;
+                      },
                       strong: ({children}) => <strong className="font-bold">{children}</strong>,
                       blockquote: ({children}) => <blockquote className="border-l-2 border-[var(--color-primary)] pl-4 my-3 italic text-[var(--color-text-muted)]">{children}</blockquote>,
                       code: ({children}) => <code className="bg-[var(--color-bg-hover)] px-1.5 py-0.5 rounded text-xs">{children}</code>,

sundol-frontend/src/app/notes/[id]/page.tsx

@@ -6,6 +6,7 @@ import AuthGuard from "@/components/auth-guard";
 import NavBar from "@/components/nav-bar";
 import { useApi } from "@/lib/use-api";
 import ReactMarkdown from "react-markdown";
+import SpeakableText from "@/components/speakable-text";
 
 interface NoteDetail {
   ID: string;
@@ -183,10 +184,18 @@ export default function NoteDetailPage() {
                   h1: ({children}) => <h1 className="text-xl font-bold mt-6 mb-3">{children}</h1>,
                   h2: ({children}) => <h2 className="text-lg font-bold mt-5 mb-2">{children}</h2>,
                   h3: ({children}) => <h3 className="text-base font-bold mt-4 mb-2">{children}</h3>,
-                  p: ({children}) => <p className="mb-3">{children}</p>,
+                  p: ({children, node}) => {
+                    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                    const txt = node?.children?.map((c: any) => c.type === 'text' ? c.value : '').join('') || '';
+                    return <p className="mb-3"><SpeakableText text={txt}>{children}</SpeakableText></p>;
+                  },
                   ul: ({children}) => <ul className="list-disc ml-5 mb-3 space-y-1">{children}</ul>,
                   ol: ({children}) => <ol className="list-decimal ml-5 mb-3 space-y-1">{children}</ol>,
-                  li: ({children}) => <li className="leading-relaxed">{children}</li>,
+                  li: ({children, node}) => {
+                    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                    const txt = node?.children?.map((c: any) => c.type === 'text' ? c.value : '').join('') || '';
+                    return <li className="leading-relaxed"><SpeakableText text={txt}>{children}</SpeakableText></li>;
+                  },
                   strong: ({children}) => <strong className="font-bold">{children}</strong>,
                   blockquote: ({children}) => <blockquote className="border-l-2 border-[var(--color-primary)] pl-4 my-3 italic text-[var(--color-text-muted)]">{children}</blockquote>,
                 }}

sundol-frontend/src/app/tts/page.tsx

@@ -45,8 +45,21 @@ export default function TTSPage() {
   }, []);
 
   const fetchProfiles = () => {
+    // 캐시 먼저
+    const cached = localStorage.getItem("tts_profiles");
+    if (cached) {
+      try {
+        const data = JSON.parse(cached);
+        setProfiles(data);
+        if (data.length > 0 && !selectedProfile) setSelectedProfile(data[0].id);
+      } catch {}
+    }
     fetch("/api/tts/profiles").then(r => r.json())
-      .then(setProfiles).catch(() => {});
+      .then(data => {
+        setProfiles(data);
+        localStorage.setItem("tts_profiles", JSON.stringify(data));
+        if (data.length > 0 && !selectedProfile) setSelectedProfile(data[0].id);
+      }).catch(() => {});
   };
 
   const startRecording = async () => {
@@ -97,6 +110,7 @@ export default function TTSPage() {
       setRecordedUrl(null);
       setUploadedFile(null);
       fetchProfiles();
+      localStorage.removeItem("tts_profiles"); // 캐시 강제 갱신
       setSelectedProfile(result.id);
       setTab("generate");
     } catch (err) {
@@ -125,9 +139,10 @@ export default function TTSPage() {
       fd.append("text", text);
       fd.append("profile_id", selectedProfile);
       fd.append("language", language);
-      const res = await fetch("/api/tts/generate", { method: "POST", body: fd });
+      const res = await fetch("/api/tts/speak", { method: "POST", body: fd });
       if (!res.ok) throw new Error(`HTTP ${res.status}`);
       const blob = await res.blob();
+      if (blob.size < 100) throw new Error("Empty audio");
       setOutputUrl(URL.createObjectURL(blob));
     } catch (err) {
       setError("생성 실패: " + (err instanceof Error ? err.message : ""));

sundol-frontend/src/components/speakable-text.tsx

@@ -0,0 +1,85 @@
+"use client";
+
+import { useState, useRef, useEffect } from "react";
+
+interface SpeakableProps {
+  children: React.ReactNode;
+  text: string;
+}
+
+let cachedProfileId: string | null = null;
+let profileChecked = false;
+
+export default function SpeakableText({ children, text }: SpeakableProps) {
+  const [playing, setPlaying] = useState(false);
+  const [loading, setLoading] = useState(false);
+  const [hasProfile, setHasProfile] = useState(false);
+  const audioRef = useRef<HTMLAudioElement | null>(null);
+
+  useEffect(() => {
+    if (profileChecked) {
+      setHasProfile(!!cachedProfileId);
+      return;
+    }
+    try {
+      const profiles = JSON.parse(localStorage.getItem("tts_profiles") || "[]");
+      if (profiles.length > 0) {
+        cachedProfileId = profiles[0].id;
+        setHasProfile(true);
+      }
+      profileChecked = true;
+    } catch {}
+  }, []);
+
+  const handleSpeak = async (e: React.MouseEvent) => {
+    e.preventDefault();
+    e.stopPropagation();
+
+    if (playing) {
+      audioRef.current?.pause();
+      setPlaying(false);
+      return;
+    }
+
+    if (!cachedProfileId || text.length < 5) return;
+
+    setLoading(true);
+    try {
+      const fd = new FormData();
+      fd.append("text", text);
+      fd.append("profile_id", cachedProfileId);
+      fd.append("language", "Korean");
+      const res = await fetch("/api/tts/speak", { method: "POST", body: fd });
+      if (!res.ok) { setLoading(false); return; }
+      const blob = await res.blob();
+      if (blob.size < 200) { setLoading(false); return; }
+
+      const url = URL.createObjectURL(blob);
+      const audio = new Audio(url);
+      audioRef.current = audio;
+      audio.onended = () => setPlaying(false);
+      setPlaying(true);
+      setLoading(false);
+      audio.play();
+    } catch {
+      setLoading(false);
+    }
+  };
+
+  if (!hasProfile || text.length < 5) return <>{children}</>;
+
+  return (
+    <>
+      {children}
+      <button
+        onClick={handleSpeak}
+        disabled={loading}
+        className="inline-flex items-center ml-1 text-[var(--color-text-muted)] hover:text-[var(--color-primary)] disabled:opacity-30 align-middle"
+        title={playing ? "중지" : "읽어주기"}
+        style={{ fontSize: "0.85em", verticalAlign: "middle", cursor: "pointer" }}
+      >
+        {loading ? "⏳" : playing ? "⏹" : "🔊"}
+      </button>
+    </>
+  );
+}

sundol-frontend/src/components/tts-reader.tsx

@@ -0,0 +1,159 @@
+"use client";
+
+import { useState, useEffect, useRef } from "react";
+
+interface TTSReaderProps {
+  text: string;
+}
+
+interface VoiceProfile {
+  id: string;
+  name: string;
+}
+
+export default function TTSReader({ text }: TTSReaderProps) {
+  const [profiles, setProfiles] = useState<VoiceProfile[]>([]);
+  const [selectedProfile, setSelectedProfile] = useState("");
+  const [generating, setGenerating] = useState(false);
+  const [playing, setPlaying] = useState(false);
+  const [progress, setProgress] = useState("");
+  const audioRef = useRef<HTMLAudioElement | null>(null);
+  const stoppedRef = useRef(false);
+  const audioUrlsRef = useRef<string[]>([]);
+
+  useEffect(() => {
+    // localStorage 캐시
+    const cached = localStorage.getItem("tts_profiles");
+    if (cached) {
+      try {
+        const data = JSON.parse(cached);
+        setProfiles(data);
+        if (data.length > 0) setSelectedProfile(data[0].id);
+      } catch {}
+    }
+    // 백그라운드에서 갱신 (블록 안 됨)
+    fetch("/api/tts/profiles").then(r => r.json()).then(data => {
+      setProfiles(data);
+      if (data.length > 0 && !selectedProfile) setSelectedProfile(data[0].id);
+      localStorage.setItem("tts_profiles", JSON.stringify(data));
+    }).catch(() => {});
+  }, []);
+
+  const toSentences = (md: string): string[] => {
+    return md
+      .replace(/^#+\s+.*$/gm, "")
+      .replace(/\*\*/g, "")
+      .replace(/^[-*]\s+/gm, "")
+      .replace(/^>\s+/gm, "")
+      .replace(/---+/g, "")
+      .replace(/\[([^\]]+)\]\([^)]+\)/g, "$1")
+      .split("\n")
+      .map(s => s.trim())
+      .filter(s => s.length >= 10);
+  };
+
+  // 직접 동기 호출 — 바로 wav 반환
+  const speak = async (chunk: string): Promise<string | null> => {
+    const fd = new FormData();
+    fd.append("text", chunk);
+    fd.append("profile_id", selectedProfile);
+    fd.append("language", "Korean");
+    const res = await fetch("/api/tts/speak", { method: "POST", body: fd });
+    if (!res.ok) return null;
+    const blob = await res.blob();
+    return blob.size > 100 ? URL.createObjectURL(blob) : null;
+  };
+
+  const handleGenerate = async () => {
+    if (!selectedProfile || !text.trim()) return;
+    setGenerating(true);
+    setPlaying(true);
+    stoppedRef.current = false;
+    audioUrlsRef.current = [];
+
+    const sentences = toSentences(text);
+    let isAudioPlaying = false;
+    let playIdx = 0;
+
+    const playNext = () => {
+      if (stoppedRef.current) return;
+      if (playIdx >= audioUrlsRef.current.length) { isAudioPlaying = false; return; }
+      isAudioPlaying = true;
+      const a = new Audio(audioUrlsRef.current[playIdx++]);
+      audioRef.current = a;
+      a.onended = () => {
+        if (stoppedRef.current) return;
+        playIdx < audioUrlsRef.current.length ? playNext() : (isAudioPlaying = false);
+      };
+      a.play();
+    };
+
+    for (let i = 0; i < sentences.length; i++) {
+      if (stoppedRef.current) break;
+      setProgress(`${i + 1}/${sentences.length}`);
+      const url = await speak(sentences[i]);
+      if (url && !stoppedRef.current) {
+        audioUrlsRef.current.push(url);
+        if (!isAudioPlaying) playNext();
+      }
+    }
+
+    setGenerating(false);
+    setProgress("");
+    if (!isAudioPlaying) setPlaying(false);
+  };
+
+  const handleStop = () => {
+    stoppedRef.current = true;
+    audioRef.current?.pause();
+    setPlaying(false);
+    setGenerating(false);
+    setProgress("");
+  };
+
+  const handleReplay = () => {
+    if (audioUrlsRef.current.length === 0) return;
+    stoppedRef.current = false;
+    setPlaying(true);
+    let idx = 0;
+    const play = () => {
+      if (idx >= audioUrlsRef.current.length || stoppedRef.current) { setPlaying(false); return; }
+      const audio = new Audio(audioUrlsRef.current[idx]);
+      audioRef.current = audio;
+      idx++;
+      audio.onended = play;
+      audio.play();
+    };
+    play();
+  };
+
+  if (profiles.length === 0) return null;
+
+  return (
+    <div className="flex items-center gap-2 flex-wrap">
+      <select value={selectedProfile} onChange={e => setSelectedProfile(e.target.value)}
+        className="text-xs px-2 py-1 rounded bg-[var(--color-bg-hover)] border border-[var(--color-border)]">
+        {profiles.map(p => <option key={p.id} value={p.id}>{p.name}</option>)}
+      </select>
+
+      {playing || generating ? (
+        <button onClick={handleStop}
+          className="text-xs px-3 py-1 bg-red-500/20 text-red-400 rounded hover:bg-red-500/30">
+          {progress || "중지"}
+        </button>
+      ) : (
+        <button onClick={handleGenerate} disabled={!selectedProfile}
+          className="text-xs px-3 py-1 bg-[var(--color-primary)]/20 text-[var(--color-primary)] rounded hover:bg-[var(--color-primary)]/30 disabled:opacity-40">
+          읽어주기
+        </button>
+      )}
+
+      {audioUrlsRef.current.length > 0 && !playing && !generating && (
+        <button onClick={handleReplay}
+          className="text-xs px-3 py-1 bg-[var(--color-bg-hover)] border border-[var(--color-border)] rounded">
+          다시 재생
+        </button>
+      )}
+    </div>
+  );
+}

tts-server.py

@@ -1,159 +1,94 @@
 """
-Qwen3-TTS Voice Clone API Server
-별도 프로세스로 실행 (GPU 메모리 관리를 위해)
+Qwen3-TTS Voice Clone API Server (최적화 버전)
+- 0.6B 모델 사용 (A10 속도 최적화)
+- 모델 1회 로드, voice clone prompt 캐시
+- inference_mode, bf16
+- 문장 단위 분할
 """
 import os
 import io
-import base64
-import tempfile
-import torch
-import soundfile as sf
-import numpy as np
-from fastapi import FastAPI, UploadFile, File, Form
-from fastapi.responses import StreamingResponse
-from fastapi.middleware.cors import CORSMiddleware
-
 import json
 import pickle
+import re
+import tempfile
+import time
+import uuid
+import threading
+
+import numpy as np
+import soundfile as sf
+import torch
+from fastapi import FastAPI, UploadFile, File, Form
+from fastapi.responses import StreamingResponse, FileResponse
+from fastapi.middleware.cors import CORSMiddleware
 
 app = FastAPI()
 app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
 
-model = None
 PROFILES_DIR = os.path.join(os.path.dirname(__file__), "voice-profiles")
 os.makedirs(PROFILES_DIR, exist_ok=True)
 
+MODEL_NAME = "Qwen/Qwen3-TTS-12Hz-1.7B-Base"
+model = None
+prompt_cache = {}  # profile_id → voice_clone_prompt
+
+
 def get_model():
     global model
     if model is None:
         from qwen_tts import Qwen3TTSModel
-        print("Loading Qwen3-TTS model...")
+        print(f"Loading {MODEL_NAME}...")
+        torch.set_grad_enabled(False)
+        torch.backends.cuda.matmul.allow_tf32 = True
+        torch.backends.cudnn.allow_tf32 = True
+        dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
         model = Qwen3TTSModel.from_pretrained(
-            "Qwen/Qwen3-TTS-12Hz-1.7B-Base",
-            device_map="cuda:0",
-            dtype=torch.bfloat16,
+            MODEL_NAME, device_map="cuda:0", dtype=dtype,
         )
+        # 프로필 프롬프트 캐시 로드
+        load_all_prompts()
         print("Model loaded!")
     return model
 
+
+def load_all_prompts():
+    """모든 프로필의 voice clone prompt를 메모리에 캐시"""
+    global prompt_cache
+    for f in os.listdir(PROFILES_DIR):
+        if f.endswith(".pkl"):
+            pid = f.replace(".pkl", "")
+            try:
+                with open(os.path.join(PROFILES_DIR, f), "rb") as fh:
+                    prompt_cache[pid] = pickle.load(fh)
+                print(f"  Cached prompt: {pid}")
+            except Exception as e:
+                print(f"  Failed to cache {pid}: {e}")
+
+
+def get_prompt(profile_id: str):
+    """캐시에서 프롬프트 가져오기, 없으면 파일에서 로드"""
+    if profile_id in prompt_cache:
+        return prompt_cache[profile_id]
+
+    pkl_path = os.path.join(PROFILES_DIR, f"{profile_id}.pkl")
+    if os.path.exists(pkl_path):
+        with open(pkl_path, "rb") as f:
+            prompt = pickle.load(f)
+        prompt_cache[profile_id] = prompt
+        return prompt
+    return None
+
+
+# === API ===
+
 @app.get("/health")
 @app.get("/api/tts/health")
 def health():
-    return {"status": "ok", "model_loaded": model is not None}
+    return {"status": "ok", "model": MODEL_NAME, "model_loaded": model is not None}
 
-@app.post("/api/tts/clone")
-async def voice_clone(
-    text: str = Form(...),
-    language: str = Form("korean"),
-    ref_audio: UploadFile = File(...),
-    ref_text: str = Form(""),
-):
-    """참조 음성으로 보이스 클로닝"""
-    m = get_model()
-
-    # 참조 음성 저장
-    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
-        content = await ref_audio.read()
-        tmp.write(content)
-        tmp_path = tmp.name
-
-    try:
-        # wav 변환 (필요 시)
-        if not ref_audio.filename.endswith(".wav"):
-            wav_path = tmp_path + "_converted.wav"
-            os.system(f'ffmpeg -i "{tmp_path}" -ar 16000 -ac 1 -y "{wav_path}" 2>/dev/null')
-            os.unlink(tmp_path)
-            tmp_path = wav_path
-
-        kwargs = {
-            "text": text,
-            "language": language,
-            "ref_audio": tmp_path,
-        }
-        if ref_text and ref_text.strip():
-            kwargs["ref_text"] = ref_text
-        else:
-            kwargs["x_vector_only_mode"] = True
-
-        wavs, sr = m.generate_voice_clone(**kwargs)
-        print(f"Clone generated: wavs={len(wavs)}, samples={len(wavs[0]) if len(wavs) > 0 else 0}, sr={sr}")
-
-        audio_data = np.array(wavs[0], dtype=np.float32)
-        buf = io.BytesIO()
-        sf.write(buf, audio_data, sr, format="WAV")
-        buf.seek(0)
-
-        return StreamingResponse(buf, media_type="audio/wav",
-                                 headers={"Content-Disposition": "attachment; filename=tts_output.wav"})
-    finally:
-        if os.path.exists(tmp_path):
-            os.unlink(tmp_path)
-
-@app.post("/api/tts/design")
-async def voice_design(
-    text: str = Form(...),
-    language: str = Form("korean"),
-    instruct: str = Form("A calm, professional Korean male voice"),
-):
-    """음성 디자인으로 생성 (참조 음성 없이)"""
-    m = get_model()
-    wavs, sr = m.generate_voice_design(text=text, instruct=instruct, language=language)
-
-    buf = io.BytesIO()
-    sf.write(buf, wavs[0], sr, format="WAV")
-    buf.seek(0)
-
-    return StreamingResponse(buf, media_type="audio/wav",
-                             headers={"Content-Disposition": "attachment; filename=tts_output.wav"})
-
-@app.post("/api/tts/profiles")
-async def create_profile(
-    name: str = Form(...),
-    ref_audio: UploadFile = File(...),
-    ref_text: str = Form(""),
-):
-    """음성 프로필 등록: 참조 음성으로 보이스 프로필 생성 후 저장"""
-    m = get_model()
-
-    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
-        content = await ref_audio.read()
-        tmp.write(content)
-        tmp_path = tmp.name
-
-    try:
-        if not ref_audio.filename.endswith(".wav"):
-            wav_path = tmp_path + "_converted.wav"
-            os.system(f'ffmpeg -i "{tmp_path}" -ar 16000 -ac 1 -y "{wav_path}" 2>/dev/null')
-            os.unlink(tmp_path)
-            tmp_path = wav_path
-
-        # 프로필 생성
-        kwargs = {"ref_audio": tmp_path}
-        if ref_text and ref_text.strip():
-            kwargs["ref_text"] = ref_text
-            prompt = m.create_voice_clone_prompt(**kwargs)
-        else:
-            kwargs["x_vector_only_mode"] = True
-            prompt = m.create_voice_clone_prompt(**kwargs)
-
-        # 저장
-        profile_id = name.replace(" ", "_").lower()
-        profile_path = os.path.join(PROFILES_DIR, f"{profile_id}.pkl")
-        meta_path = os.path.join(PROFILES_DIR, f"{profile_id}.json")
-
-        with open(profile_path, "wb") as f:
-            pickle.dump(prompt, f)
-        with open(meta_path, "w") as f:
-            json.dump({"id": profile_id, "name": name, "ref_text": ref_text}, f, ensure_ascii=False)
-
-        return {"id": profile_id, "name": name, "status": "created"}
-    finally:
-        if os.path.exists(tmp_path):
-            os.unlink(tmp_path)
 
 @app.get("/api/tts/profiles")
 def list_profiles():
-    """저장된 음성 프로필 목록"""
     profiles = []
     for f in os.listdir(PROFILES_DIR):
         if f.endswith(".json"):
@@ -161,49 +96,119 @@ def list_profiles():
                 profiles.append(json.load(fh))
     return profiles
 
+
+@app.post("/api/tts/profiles")
+async def create_profile(
+    name: str = Form(...),
+    ref_audio: UploadFile = File(...),
+    ref_text: str = Form(""),
+):
+    m = get_model()
+    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
+        content = await ref_audio.read()
+        tmp.write(content)
+        tmp_path = tmp.name
+
+    try:
+        if not ref_audio.filename.endswith(".wav"):
+            wav_path = tmp_path + "_converted.wav"
+            os.system(f'ffmpeg -i "{tmp_path}" -ar 16000 -ac 1 -y "{wav_path}" 2>/dev/null')
+            os.unlink(tmp_path)
+            tmp_path = wav_path
+
+        kwargs = {"ref_audio": tmp_path}
+        if ref_text and ref_text.strip():
+            kwargs["ref_text"] = ref_text
+        else:
+            kwargs["x_vector_only_mode"] = True
+
+        with torch.inference_mode():
+            prompt = m.create_voice_clone_prompt(**kwargs)
+
+        profile_id = name.replace(" ", "_").lower()
+
+        # wav, pkl, json 저장
+        import shutil
+        shutil.copy2(tmp_path, os.path.join(PROFILES_DIR, f"{profile_id}.wav"))
+        with open(os.path.join(PROFILES_DIR, f"{profile_id}.pkl"), "wb") as f:
+            pickle.dump(prompt, f)
+        with open(os.path.join(PROFILES_DIR, f"{profile_id}.json"), "w") as f:
+            json.dump({"id": profile_id, "name": name, "ref_text": ref_text}, f, ensure_ascii=False)
+
+        # 캐시에 추가
+        prompt_cache[profile_id] = prompt
+
+        return {"id": profile_id, "name": name, "status": "created"}
+    finally:
+        if os.path.exists(tmp_path):
+            os.unlink(tmp_path)
+
+
 @app.delete("/api/tts/profiles/{profile_id}")
 def delete_profile(profile_id: str):
-    """음성 프로필 삭제"""
-    pkl = os.path.join(PROFILES_DIR, f"{profile_id}.pkl")
-    meta = os.path.join(PROFILES_DIR, f"{profile_id}.json")
-    if os.path.exists(pkl): os.unlink(pkl)
-    if os.path.exists(meta): os.unlink(meta)
+    for ext in [".pkl", ".json", ".wav"]:
+        p = os.path.join(PROFILES_DIR, f"{profile_id}{ext}")
+        if os.path.exists(p):
+            os.unlink(p)
+    prompt_cache.pop(profile_id, None)
     return {"status": "deleted"}
 
-@app.post("/api/tts/generate")
-async def generate_from_profile(
+
+@app.post("/api/tts/speak")
+async def speak(
     text: str = Form(...),
     profile_id: str = Form(...),
-    language: str = Form("korean"),
+    language: str = Form("Korean"),
 ):
-    """저장된 음성 프로필로 TTS 생성"""
+    """한 문장 TTS — 캐시된 프롬프트 사용, 바로 wav 반환"""
     m = get_model()
 
-    profile_path = os.path.join(PROFILES_DIR, f"{profile_id}.pkl")
-    if not os.path.exists(profile_path):
-        return {"error": f"Profile '{profile_id}' not found"}, 404
+    prompt = get_prompt(profile_id)
+    if prompt is None:
+        # 프롬프트가 없으면 ref_audio로 직접
+        meta_path = os.path.join(PROFILES_DIR, f"{profile_id}.json")
+        ref_audio_path = os.path.join(PROFILES_DIR, f"{profile_id}.wav")
+        if not os.path.exists(ref_audio_path):
+            return {"error": "Profile not found"}, 404
 
-    with open(profile_path, "rb") as f:
-        prompt = pickle.load(f)
+        with open(meta_path) as f:
+            meta = json.load(f)
 
-    print(f"Generating with profile '{profile_id}', text='{text[:50]}...', language={language}")
-    wavs, sr = m.generate_voice_clone(
-        text=text,
-        language=language,
-        voice_clone_prompt=prompt,
-    )
-    print(f"Generated: wavs={len(wavs)}, samples={len(wavs[0]) if len(wavs) > 0 else 0}, sr={sr}")
+        kwargs = {"text": text, "language": language, "ref_audio": ref_audio_path}
+        if meta.get("ref_text"):
+            kwargs["ref_text"] = meta["ref_text"]
+        else:
+            kwargs["x_vector_only_mode"] = True
 
-    if len(wavs) == 0 or len(wavs[0]) == 0:
-        return {"error": "Empty audio generated"}, 500
+        start = time.perf_counter()
+        with torch.inference_mode():
+            wavs, sr = m.generate_voice_clone(**kwargs)
+        elapsed = time.perf_counter() - start
+    else:
+        start = time.perf_counter()
+        with torch.inference_mode():
+            wavs, sr = m.generate_voice_clone(
+                text=text, language=language, voice_clone_prompt=prompt,
+            )
+        elapsed = time.perf_counter() - start
 
     audio_data = np.array(wavs[0], dtype=np.float32)
+    print(f"speak: {len(text)} chars → {len(audio_data)/sr:.1f}s audio in {elapsed:.1f}s")
+
     buf = io.BytesIO()
     sf.write(buf, audio_data, sr, format="WAV")
     buf.seek(0)
+    return StreamingResponse(buf, media_type="audio/wav")
+
+
+from fastapi.responses import JSONResponse
+
+@app.exception_handler(Exception)
+async def global_exception_handler(request, exc):
+    import traceback
+    traceback.print_exc()
+    return JSONResponse(status_code=500, content={"error": str(exc)})
 
-    return StreamingResponse(buf, media_type="audio/wav",
-                             headers={"Content-Disposition": "attachment; filename=tts_output.wav"})
 
 if __name__ == "__main__":
     import uvicorn