TTS Coqui Developer Experience DX

Coqui TTS Developer Experience

Coqui TTS Text-to-Speech Python Deep Learning VITS Tacotron2 Voice Cloning Fine-tune API Server ONNX Production Multi-speaker

Quick Start

# === Coqui TTS Quick Start ===

# pip install TTS
# pip install TTS[server]  # สำหรับ Demo Server

# CLI Usage
# tts --list_models  # ดู Models ทั้งหมด
# tts --text "Hello World" \
#     --model_name tts_models/en/ljspeech/vits \
#     --out_path output.wav
#
# # Multi-speaker
# tts --text "Hello" \
#     --model_name tts_models/en/vctk/vits \
#     --speaker_idx "p225" \
#     --out_path output.wav
#
# # Voice Cloning
# tts --text "Hello" \
#     --model_name tts_models/multilingual/multi-dataset/your_tts \
#     --speaker_wav reference.wav \
#     --language_idx "en" \
#     --out_path cloned.wav

# Python API
# from TTS.api import TTS
#
# # Single Speaker
# tts = TTS("tts_models/en/ljspeech/vits", gpu=True)
# tts.tts_to_file(text="Hello World", file_path="output.wav")
#
# # Multi Speaker
# tts = TTS("tts_models/en/vctk/vits", gpu=True)
# tts.tts_to_file(text="Hello", file_path="out.wav", speaker="p225")
#
# # Voice Cloning
# tts = TTS("tts_models/multilingual/multi-dataset/your_tts", gpu=True)
# tts.tts_to_file(text="Hello", file_path="cloned.wav",
#                 speaker_wav="reference.wav", language="en")

from dataclasses import dataclass

@dataclass
class TTSModel:
    name: str
    model_id: str
    languages: str
    speakers: str
    quality: str

models = [
    TTSModel("VITS (LJSpeech)",
        "tts_models/en/ljspeech/vits",
        "English",
        "Single",
        "สูงมาก End-to-end Fast"),
    TTSModel("VITS (VCTK)",
        "tts_models/en/vctk/vits",
        "English",
        "Multi (109 speakers)",
        "สูง Multi-speaker"),
    TTSModel("YourTTS",
        "tts_models/multilingual/multi-dataset/your_tts",
        "English Portuguese French",
        "Voice Cloning",
        "สูง Zero-shot Cloning"),
    TTSModel("Tacotron2-DDC",
        "tts_models/en/ljspeech/tacotron2-DDC",
        "English",
        "Single",
        "สูง Classic Architecture"),
]

print("=== Available Models ===")
for m in models:
    print(f"  [{m.name}]")
    print(f"    ID: {m.model_id}")
    print(f"    Lang: {m.languages} | Speakers: {m.speakers}")
    print(f"    Quality: {m.quality}")

Fine-tuning

# === Fine-tune Coqui TTS ===

# Dataset Structure
# dataset/
#   wavs/
#     audio_001.wav  (22050Hz Mono 16-bit)
#     audio_002.wav
#   metadata.csv
#     audio_001|Hello this is a test
#     audio_002|Another sentence here

# Training Config (config.json)
# {
#   "model": "vits",
#   "audio": {"sample_rate": 22050},
#   "datasets": [{
#     "formatter": "ljspeech",
#     "path": "./dataset/",
#     "meta_file_train": "metadata.csv"
#   }],
#   "training": {
#     "batch_size": 32,
#     "epochs": 1000,
#     "lr": 0.0002,
#     "print_step": 50,
#     "save_step": 1000
#   }
# }

# python -m TTS.bin.train_tts --config_path config.json

@dataclass
class TrainingTip:
    aspect: str
    recommendation: str
    common_mistake: str
    fix: str

tips = [
    TrainingTip("Dataset Size",
        "อย่างน้อย 2-5 ชั่วโมง Audio สำหรับ Fine-tune",
        "Audio น้อยเกินไป Model ไม่เรียนรู้",
        "ใช้ Data Augmentation หรือหา Audio เพิ่ม"),
    TrainingTip("Audio Quality",
        "22050Hz Mono 16-bit WAV ไม่มี Noise Background",
        "Audio มี Noise Model เรียนรู้ Noise ด้วย",
        "ใช้ Noise Reduction ก่อน Train เช่น RNNoise"),
    TrainingTip("Transcription",
        "ต้องตรงกับ Audio 100% Punctuation ถูกต้อง",
        "Text ไม่ตรง Audio Model สับสน",
        "ใช้ Whisper Transcribe แล้ว Manual Review"),
    TrainingTip("Learning Rate",
        "เริ่ม 0.0002 สำหรับ Fine-tune",
        "LR สูงเกินไป Model Diverge",
        "ใช้ LR Scheduler ลด LR เมื่อ Loss ไม่ลด"),
    TrainingTip("Batch Size",
        "32 (GPU 16GB) ลดถ้า OOM",
        "Batch Size ใหญ่เกิน GPU Memory",
        "ลด Batch Size หรือใช้ Gradient Accumulation"),
]

print("=== Training Tips ===")
for t in tips:
    print(f"\n  [{t.aspect}] {t.recommendation}")
    print(f"    Mistake: {t.common_mistake}")
    print(f"    Fix: {t.fix}")

Production API

# === Production TTS API ===

# # FastAPI TTS Server
# from fastapi import FastAPI, Response
# from TTS.api import TTS
# import io, soundfile as sf
#
# app = FastAPI()
# tts = TTS("tts_models/en/ljspeech/vits", gpu=True)
#
# @app.post("/tts")
# async def text_to_speech(text: str, speaker: str = None):
#     wav = tts.tts(text=text, speaker=speaker)
#     buffer = io.BytesIO()
#     sf.write(buffer, wav, 22050, format="WAV")
#     buffer.seek(0)
#     return Response(content=buffer.read(),
#                     media_type="audio/wav")
#
# # gunicorn main:app -w 1 -k uvicorn.workers.UvicornWorker
# # (1 worker per GPU)

@dataclass
class ProductionConfig:
    component: str
    config: str
    purpose: str
    scaling: str

production = [
    ProductionConfig("FastAPI + Gunicorn",
        "1 Worker per GPU + Uvicorn",
        "Production HTTP API Server",
        "Horizontal: เพิ่ม Replicas ตาม GPU"),
    ProductionConfig("ONNX Runtime",
        "Export Model → ONNX → ONNXRuntime Inference",
        "Inference เร็วขึ้น 2-5x ลด Memory",
        "รองรับ CPU Inference ไม่ต้อง GPU"),
    ProductionConfig("Redis Queue",
        "Celery + Redis สำหรับ Async TTS",
        "Queue Request ไม่ Block API",
        "เพิ่ม Worker ตาม Backlog"),
    ProductionConfig("Audio Cache",
        "Redis/S3 Cache Generated Audio",
        "ลด Compute สำหรับ Text ซ้ำ",
        "Cache Hit Rate 30-50% ลด Cost"),
    ProductionConfig("Nginx + Rate Limit",
        "Nginx Reverse Proxy + limit_req",
        "ป้องกัน Abuse DDoS",
        "10 req/s per IP"),
    ProductionConfig("Monitoring",
        "Prometheus + Grafana",
        "Latency Request Rate GPU Utilization",
        "Alert เมื่อ Latency > 5s หรือ GPU > 90%"),
]

print("=== Production Config ===")
for p in production:
    print(f"  [{p.component}] {p.config}")
    print(f"    Purpose: {p.purpose}")
    print(f"    Scaling: {p.scaling}")

เคล็ดลับ

• VITS: ใช้ VITS Model สำหรับ Production ดีที่สุด End-to-end
• ONNX: Export ONNX Inference เร็วขึ้น 2-5x รองรับ CPU
• Cache: Cache Audio ที่ Generate แล้ว ลด Compute 30-50%
• GPU: 1 Worker per GPU ไม่ Share GPU ระหว่าง Workers
• Dataset: Audio Quality สำคัญที่สุด ใช้ Noise Reduction ก่อน Train

Coqui TTS คืออะไร

Open Source TTS Python Deep Learning VITS Tacotron2 Multi-speaker Voice Cloning Fine-tune API Server ONNX Pre-trained Models MPL

สรุป

Coqui TTS Developer Experience VITS Tacotron2 Voice Cloning Fine-tune FastAPI ONNX Redis Cache GPU Monitoring Production