---
id: wiki-2026-0508-speech-synthesis
title: Speech Synthesis
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [tts, text-to-speech, voice-synthesis, neural-tts]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [tts, audio, ai, voice-cloning, generative]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: elevenlabs-openai-coqui
---

# Speech Synthesis

## 매 한 줄
> **"매 text → audio waveform — 매 neural acoustic + vocoder pipeline"**. Speech synthesis (TTS) 는 매 text 를 자연스러운 speech 로 변환 — 매 2026 의 ElevenLabs v3 / OpenAI gpt-4o-tts / Coqui XTTS-v2 가 human-indistinguishable level. 매 voice cloning, emotion control, multi-lingual 의 production-ready.

## 매 핵심

### 매 진화 단계
- **Concatenative (1990s)**: phoneme database 의 splice — 매 robotic.
- **Parametric / HMM (2000s)**: statistical 의 generation — 매 muffled.
- **Neural (2017+)**: Tacotron 2 + WaveNet — 매 인간 수준 prosody.
- **End-to-end (2021+)**: VITS, NaturalSpeech — 매 single model.
- **Foundation TTS (2024+)**: Voicebox, NaturalSpeech 3, ElevenLabs v3 — 매 zero-shot voice cloning, emotion, style.

### 매 Architecture
- **Acoustic model**: text → mel-spectrogram (Tacotron, FastSpeech 2, VITS).
- **Vocoder**: mel → waveform (WaveNet, HiFi-GAN, BigVGAN).
- **End-to-end**: text → waveform 직접 (VITS, NaturalSpeech).
- **LLM-based (2024+)**: AudioLM, Voicebox — 매 token-based audio LM.

### 매 응용
1. **Audiobook / podcast** — 매 ElevenLabs Studio.
2. **Voice agent** — 매 real-time conversational AI (Vapi, Retell).
3. **Game NPC** — 매 dynamic dialog (AI Dungeon, Inworld).
4. **Accessibility** — 매 screen reader, dyslexia aid.
5. **Localization** — 매 video dubbing (HeyGen, ElevenLabs dubbing).

## 💻 패턴

### ElevenLabs API (production default)
```python
from elevenlabs import ElevenLabs

client = ElevenLabs(api_key="...")
audio = client.text_to_speech.convert(
    voice_id="21m00Tcm4TlvDq8ikWAM",  # Rachel
    model_id="eleven_v3",
    text="안녕하세요, 매 wiki cleanup batch.",
    voice_settings={"stability": 0.5, "similarity_boost": 0.75},
)
with open("out.mp3", "wb") as f:
    for chunk in audio: f.write(chunk)
```

### OpenAI TTS (GPT-5 era)
```python
from openai import OpenAI
client = OpenAI()
resp = client.audio.speech.create(
    model="gpt-4o-mini-tts",
    voice="nova",
    input="Hello world",
    instructions="Speak in a calm, encouraging tone.",
)
resp.stream_to_file("out.mp3")
```

### Coqui XTTS-v2 (open-source, voice cloning)
```python
from TTS.api import TTS
tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cuda")
tts.tts_to_file(
    text="복제된 voice 의 sample",
    speaker_wav="reference_6sec.wav",  # 6s reference 의 zero-shot clone
    language="ko",
    file_path="cloned.wav",
)
```

### Streaming (low-latency agent)
```python
from elevenlabs import stream
audio_stream = client.text_to_speech.convert_as_stream(
    voice_id="...", model_id="eleven_flash_v2_5",  # ~75ms TTFB
    text="streamed reply",
)
stream(audio_stream)
```

### Real-time (WebSocket, sub-200ms)
```python
import websockets, json, asyncio
async def speak():
    uri = "wss://api.elevenlabs.io/v1/text-to-speech/{vid}/stream-input"
    async with websockets.connect(uri, extra_headers={"xi-api-key": KEY}) as ws:
        await ws.send(json.dumps({"text": " ", "voice_settings": {...}}))
        for chunk in llm_stream():  # token-by-token from LLM
            await ws.send(json.dumps({"text": chunk}))
        await ws.send(json.dumps({"text": ""}))
```

### MLX local TTS (Apple Silicon)
```python
from mlx_audio.tts.generate import generate_audio
generate_audio(text="local synthesis", model_path="mlx-community/Kokoro-82M-bf16",
               voice="af_heart", output_path="local.wav")
```

### SSML / prosody control
```xml
<speak>
  <prosody rate="slow" pitch="+10%">매 천천히, 높은 톤</prosody>
  <break time="500ms"/>
  <emphasis level="strong">중요</emphasis>합니다.
</speak>
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Production voice agent | ElevenLabs Flash v2.5 (75ms TTFB) |
| Highest quality narration | ElevenLabs v3 / OpenAI gpt-4o-tts |
| Voice cloning (legal consent) | XTTS-v2 (open) / ElevenLabs Pro |
| On-device / privacy | Kokoro-82M (MLX) / Piper |
| Multilingual dub | ElevenLabs dubbing API / HeyGen |
| Cost-sensitive batch | OpenAI tts-1 / self-hosted Coqui |

**기본값**: ElevenLabs Flash v2.5 (real-time) / v3 (quality batch). 매 on-device 는 Kokoro-82M.

## 🔗 Graph
- 부모: [[Generative-AI]]
- 변형: [[Voice-Cloning]]
- Adjacent: [[ASR]] · [[Whisper]] · [[Multimodal-LLM]]

## 🤖 LLM 활용
**언제**: voice agent (LLM → TTS pipeline), dynamic narration, accessibility, localization.
**언제 X**: legal/medical 의 critical announcements (human voice 의 trust 필요), sung performance (specialized 모델 사용).

## ❌ 안티패턴
- **Voice cloning 의 consent 없이**: deepfake 의 legal/ethical violation. ElevenLabs Voice Verification 사용.
- **Long-form 의 single API call**: timeout / cost spike. Chunk by sentence + stream.
- **No SSML for prosody**: monotone delivery. `<prosody>` 활용.
- **Wrong sample rate mixing**: 22kHz vs 44.1kHz mix → distortion. Resample first.
- **Pronunciation 의 무방치**: 고유명사 mispronounce. Phoneme override / lexicon 사용.

## 🧪 검증 / 중복
- Verified (ElevenLabs docs 2026, OpenAI Audio API, Coqui XTTS-v2 paper, Apple MLX-Audio).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — modern TTS landscape (ElevenLabs v3, OpenAI gpt-4o-tts, XTTS-v2, Kokoro) |