VoiceBox：开源 AI 语音工作室——声音克隆、听写与语音合成

# 从 PyPI 安装 VoiceBox
pip install voicebox-ai

# 验证安装
voicebox --version

# 初始化应用
voicebox init --model qwen3-tts

# 克隆仓库
git clone https://github.com/jamiepine/voicebox.git
cd voicebox

# 创建虚拟环境
python -m venv .venv
source .venv/bin/activate

# 安装依赖
pip install -r requirements.txt

# 以开发模式安装
pip install -e .

# 下载默认声音模型
voicebox download-models --all

# 拉取官方镜像
docker pull jamiepine/voicebox:latest

# 带 GPU 支持运行（NVIDIA）
docker run -d \
  --name voicebox \
  --gpus all \
  -p 8000:8000 \
  -v ${HOME}/voicebox-data:/data \
  -e VOICEBOX_MODEL=qwen3-tts \
  jamiepine/voicebox:latest

# 在 Apple Silicon 上运行（无需 GPU 参数）
docker run -d \
  --name voicebox \
  -p 8000:8000 \
  -v ${HOME}/voicebox-data:/data \
  -e VOICEBOX_MODEL=qwen3-tts \
  jamiepine/voicebox:latest

# 从微软商店安装 Python 3.11+
# 然后安装 VoiceBox
pip install voicebox-ai

# GPU 加速需要安装 CUDA 工具包
# 下载地址: https://developer.nvidia.com/cuda-downloads

# 初始化 VoiceBox
voicebox init --gpu cuda

# 用内置录音器录制音频
voicebox record --output sample.wav --duration 30

# 或上传现有音频文件
voicebox clone --audio my_voice_sample.mp3 --name "my-voice"

# VoiceBox 自动处理音频并提取声音特征

from voicebox.engine import VoiceCloner
from voicebox.audio import AudioProcessor

# 初始化克隆器
cloner = VoiceCloner(model="qwen3-tts-voice-clone")

# 加载并预处理参考音频
processor = AudioProcessor()
reference = processor.load_audio("sample.wav")
reference = processor.normalize(reference, target_rms=-20)
reference = processor.remove_noise(reference, method="spectral")

# 提取声音嵌入
embeddings = cloner.extract_embeddings(reference)

# 创建声音模型
voice_model = cloner.create_voice(
    embeddings=embeddings,
    name="my-voice",
    quality="high"
)

# 测试克隆的声音
output = voice_model.synthesize(
    text="Hello, this is my cloned voice speaking.",
    speed=1.0,
    emotion="neutral"
)
voice_model.save(output, "test_output.wav")

# 控制语速
voicebox synthesize --input script.txt --output speech.wav --speed 0.8

# 添加情绪变化
voicebox synthesize --input script.txt --output emotional.wav --emotion happy

# 调整音高
voicebox synthesize --input script.txt --output pitched.wav --pitch +200

# 组合多个参数
voicebox synthesize \
  --input script.txt \
  --output natural.wav \
  --speed 1.1 \
  --pitch +100 \
  --emotion confident \
  --clarity high

from voicebox.engine import VoiceManager

manager = VoiceManager()

# 列出所有克隆的声音
voices = manager.list_voices()
for v in voices:
    print(f"{v.name}: {v.quality} ({v.duration}s of training data)")

# 切换声音
manager.set_active_voice("my-voice")
output = manager.synthesize("Hello from my cloned voice!")

# 混合两个声音生成混合语音
hybrid = manager.blend_voices(
    voice_a="my-voice",
    voice_b="partner-voice",
    weight_a=0.7,
    weight_b=0.3
)
output = hybrid.synthesize("Blended voice output")

# 启用系统级听写
voicebox dictation --enable

# 选择识别模型
voicebox dictation --model whisper-large-v3

# 设置输出语言
voicebox dictation --language en

# 配置热键
voicebox dictation --hotkey "ctrl+space"

from voicebox.dictation import DictationEngine

# 初始化听写引擎
engine = DictationEngine(
    model="whisper-large-v3",
    language="auto",
    beam_size=5,
    vad_threshold=0.5
)

# 开始监听
engine.start_listening(
    hotkey="ctrl+shift+d",
    output_mode="clipboard",
    append_mode=True
)

# 处理一次听写会话
result = await engine.listen_session(
    timeout=300,           # 5 分钟会话
    silence_threshold=1.5, # 静音 1.5 秒后停止
    language="en"
)

print(f"Transcribed: {result.text}")
print(f"Confidence: {result.confidence:.2%}")
print(f"Words: {result.word_count}")

# 启用自动检测
voicebox dictation --auto-detect

# 指定支持的语言
voicebox dictation --languages en,zh,ko,ja,es,fr,de

# 设置主要语言（提高准确率）
voicebox dictation --primary-language en

# 简单的文本转语音
curl -X POST "https://your-voicebox/api/v1/tts" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Hello, this is a test of VoiceBox text-to-speech.",
    "voice": "default",
    "speed": 1.0,
    "output_format": "wav"
  }' \
  --output speech.wav

# 分块流式传输音频
curl -N -X POST "https://your-voicebox/api/v1/tts/stream" \
  -H "Content-Type: application/json" \
  -d '{"text": "This audio will stream in real-time...", "voice": "cloned-voice"}' \
  --output - | aplay

from voicebox.api import VoiceBoxClient

client = VoiceBoxClient("https://your-voicebox")

texts = [
    "First sentence for processing.",
    "Second sentence with different content.",
    "Third sentence in another voice.",
]

results = await client.tts.batch(
    texts=texts,
    voice="default",
    output_format="mp3",
    parallel_workers=4
)

for i, result in enumerate(results):
    print(f"Generated: speech_{i}.mp3 ({result.duration:.1f}s)")