# 传统 TTS 流程
Text → Tokenizer → Phonemes → Model → Audio
     ↑ 复杂的tokenizer需要额外处理

# VoxCPM2 流程  
Text → LocDiT Model → Audio
     ↑ 无需tokenizer，直接建模

# MiniCPM-4 架构特点
MiniCPMConfig {
    "dim": 2048,           # 模型维度
    "n_layers": 40,        # Transformer 层数
    "n_heads": 32,         # 注意力头数
    "vocab_size": 122_753, # 词表大小
    "max_seq_len": 4096,  # 最大序列长度
}

# 基础安装
pip install voxcpm

# GPU 加速（推荐）
pip install torch torchvision torchaudio
pip install voxcpm --extra-index-url https://download.pytorch.org/whl/cu118

# 验证安装
python -c "from voxcpm import VoxCPM; print('VoxCPM installed successfully!')"

from voxcpm import VoxCPM

# 加载模型
model = VoxCPM.from_pretrained("openbmb/VoxCPM2")

# 基础 TTS
text = "Hello, this is a test of VoxCPM2, a tokenizer-free TTS system."
audio = model.generate(text=text)

# 保存音频
model.save(audio, "output.wav")

# 或指定采样率
audio_48k = model.generate(
    text=text,
    sample_rate=48000,  # 48kHz 高质量输出
    cfg_value=2.0,     # Classifier-Free Guidance 强度
    inference_timesteps=10  # 扩散步数，越多越慢但质量越好
)

from voxcpm import VoxCPM

model = VoxCPM.from_pretrained("openbmb/VoxCPM2")

# Voice Design：自然语言描述音色
voice_description = "A warm, friendly female voice with a slight British accent"
audio = model.voice_design(
    text="Good morning! How can I assist you today?",
    voice_prompt=voice_description
)

model.save(audio, "voice_design_output.wav")

from voxcpm import VoxCPM

model = VoxCPM.from_pretrained("openbmb/VoxCPM2")

# 克隆参考音频
reference_audio = "reference_speaker.wav"
text_to_speak = "This voice will be cloned and used to speak this text."

# 基本克隆
audio = model.clone_voice(
    text=text_to_speak,
    ref_audio=reference_audio
)

# 带风格控制的克隆
audio_with_style = model.clone_voice(
    text=text_to_speak,
    ref_audio=reference_audio,
    style="happy",  # happy, sad, calm, excited
    speaking_rate=1.1  # 语速控制，1.0 为正常
)

model.save(audio_with_style, "cloned_voice.wav")

# Ultimate Cloning：配合转录文本，完美复刻
audio = model.ultimate_clone(
    text="Transcript of the reference audio",
    ref_audio="reference_with_transcript.wav"
)

from voxcpm import VoxCPM

model = VoxCPM.from_pretrained("openbmb/VoxCPM2")

# 批量文本合成
texts = [
    "First sentence to synthesize.",
    "Second sentence for batch processing.",
    "Third and final sentence in this batch."
]

# 批量生成
audios = model.batch_generate(
    texts=texts,
    cfg_value=2.0,
    inference_timesteps=10
)

# 保存每个音频
for i, audio in enumerate(audios):
    model.save(audio, f"batch_output_{i}.wav")

# 中文语音合成
zh_audio = model.generate(
    text="你好，这是一个中文语音合成的例子。",
    language="zh"  # 自动检测或手动指定
)

# 英文语音合成
en_audio = model.generate(
    text="This is an English speech synthesis example.",
    language="en"
)

# 日文语音合成
ja_audio = model.generate(
    text="これは日本語の音声合成の例です。",
    language="ja"
)

# 自动检测语言
auto_audio = model.generate(
    text="This text will be automatically detected.",
    language="auto"  # 自动检测
)

# 不同采样率对比
audio_16k = model.generate(text, sample_rate=16000)  # 轻量
audio_44k = model.generate(text, sample_rate=44100)  # CD音质
audio_48k = model.generate(text, sample_rate=48000)  # 母带级

# 低 CFG：更多变化，可能不太稳定
audio_var = model.generate(text, cfg_value=1.0)

# 适中 CFG：平衡质量与变化
audio_balanced = model.generate(text, cfg_value=2.0)

# 高 CFG：更稳定但可能过于平滑
audio_stable = model.generate(text, cfg_value=3.0)

# 快速推理（质量较低）
audio_fast = model.generate(text, inference_timesteps=5)

# 标准质量
audio_std = model.generate(text, inference_timesteps=10)

# 高质量（较慢）
audio_hq = model.generate(text, inference_timesteps=20)

# 极致质量
audio_max = model.generate(text, inference_timesteps=50)

from voxcpm import VoxCPM, VoxCPMFineTuner

# 加载预训练模型
model = VoxCPM.from_pretrained("openbmb/VoxCPM2")

# 配置 LoRA 微调
finetuner = VoxCPMFineTuner(model)

# 准备微调数据
train_data = [
    {"text": "训练文本1", "audio": "path/to/audio1.wav"},
    {"text": "训练文本2", "audio": "path/to/audio2.wav"},
]

# 开始微调
finetuner.finetune(
    train_data=train_data,
    lora_rank=16,           # LoRA rank
    learning_rate=1e-4,
    epochs=3,
    batch_size=2
)

# 保存微调后的模型
finetuner.save("my_finetuned_voxcpm2")

# 全量参数微调（需要更多显存）
finetuner.finetune(
    train_data=train_data,
    finetune_mode="full",  # 全量微调
    learning_rate=5e-5,
    epochs=5,
    batch_size=1,  # 全量需要更小 batch
    gradient_accumulation=4
)

# 数据格式示例
dataset = [
    {
        "text": "The quick brown fox jumps over the lazy dog.",
        "audio_path": "/path/to/audio.wav",
        "duration": 3.5,  # 秒
        "language": "en",
        "speaker_id": "speaker_001"
    },
    # ... 更多样本
]

# 保存为 JSONL 格式
import json
with open("voxcpm_ft_data.jsonl", "w") as f:
    for item in dataset:
        f.write(json.dumps(item) + "\n")

# 拉取镜像
docker pull openbmb/voxcpm2:latest

# 运行容器
docker run -p 8080:8080 \
    -v $(pwd)/models:/app/models \
    openbmb/voxcpm2:latest

# API 调用
curl -X POST http://localhost:8080/generate \
    -H "Content-Type: application/json" \
    -d '{"text": "Hello world", "language": "en"}' \
    --output output.wav

# 使用 Nano-vllm 加速推理
from voxcpm import VoxCPM

# vLLM 后端
model = VoxCPM.from_pretrained(
    "openbmb/VoxCPM2",
    backend="vllm",  # 使用 vLLM 加速
    tensor_parallel_size=2  # 多卡并行
)

# 批量推理加速
audios = model.batch_generate(texts, batch_size=32)

# macOS MPS 加速
model = VoxCPM.from_pretrained(
    "openbmb/VoxCPM2",
    device="mps"  # Apple Silicon GPU
)

audio = model.generate(text)

# 导出为 ONNX 格式
from voxcpm import export

export.to_onnx(
    model=model,
    output_path="voxcpm2.onnx",
    optimize=True  # 优化模型
)

# 安装
cargo install voxcpm

# 命令行使用
voxcpm generate --text "Hello" --output hello.wav

# 作为库使用
voxcpm-cli synthesize \
    --model openbmb/VoxCPM2 \
    --text "你好世界" \
    --language zh \
    --output chinese.wav

from voxcpm_onnxruntime import VoxCPM_ONNX

model = VoxCPM_ONNX.from_pretrained(
    "openbmb/VoxCPM2-onnx",
    providers=["CPU", "CUDA", "TensorRT"]
)

// Swift 调用
import VoxCPM_ANE

let model = VoxCPM_ANE(modelPath: "VoxCPM2.ane")
let audio = try model.generate(
    text: "Hello from Apple Neural Engine!",
    language: .english
)

# 解决方案 1: 减少 batch size
model.generate(texts, batch_size=1)

# 解决方案 2: 使用量化版本
model = VoxCPM.from_pretrained(
    "openbmb/VoxCPM2-quantized-int4"
)

# 解决方案 3: 使用 CPU
model = VoxCPM.from_pretrained(
    "openbmb/VoxCPM2",
    device="cpu"
)

# 方案 1: 增加推理步数
audio = model.generate(text, inference_timesteps=20)

# 方案 2: 调整 CFG 值
audio = model.generate(text, cfg_value=2.5)

# 方案 3: 使用高质量采样率
audio = model.generate(text, sample_rate=48000)

# 手动指定语言
audio = model.generate(
    text="这是一段中文文本",
    language="zh"  # 明确指定
)

# 支持的语言代码
LANGUAGES = ["zh", "en", "ja", "ko", "fr", "de", "es", "it", "ru", ...]