1
from voxcpm import VoxCPM
2
model = VoxCPM.from_pretrained("openbmb/VoxCPM2", load_denoiser=False)

1
from modelscope import snapshot_download
2
snapshot_download("OpenBMB/VoxCPM2", local_dir="./pretrained_models/VoxCPM2")
3

4
# 从本地加载
5
from voxcpm import VoxCPM
6
model = VoxCPM.from_pretrained("./pretrained_models/VoxCPM2", load_denoiser=False)

1
from voxcpm import VoxCPM
2
import soundfile as sf
3

4
model = VoxCPM.from_pretrained("openbmb/VoxCPM2", load_denoiser=False)
5

6
wav = model.generate(
7
    text="VoxCPM2 是目前推荐的多语言逼真语音合成版本。",
8
    cfg_value=2.0,           # CFG 引导强度
9
    inference_timesteps=10,  # 扩散步数
10
)
11
sf.write("demo.wav", wav, model.tts_model.sample_rate)

1
wav = model.generate(
2
    text="(一位年轻女性，温柔甜美的声音)你好，欢迎使用 VoxCPM2！",
3
    cfg_value=2.0,
4
    inference_timesteps=10,
5
)
6
sf.write("voice_design.wav", wav, model.tts_model.sample_rate)

1
# 基础克隆
2
wav = model.generate(
3
    text="这是 VoxCPM2 生成的克隆语音。",
4
    reference_wav_path="speaker.wav",
5
)
6

7
# 克隆 + 风格控制
8
wav = model.generate(
9
    text="(语速稍快，欢快语调)这是带风格控制的克隆语音。",
10
    reference_wav_path="speaker.wav",
11
    cfg_value=2.0,
12
    inference_timesteps=10,
13
)
14
sf.write("clone.wav", wav, model.tts_model.sample_rate)

1
wav = model.generate(
2
    text="这是 VoxCPM2 的终极克隆演示。",
3
    prompt_wav_path="speaker.wav",
4
    prompt_text="参考音频对应的文本转录内容。",
5
    reference_wav_path="speaker.wav",  # 可选，提升相似度
6
)
7
sf.write("hifi_clone.wav", wav, model.tts_model.sample_rate)

1
import numpy as np
2

3
chunks = []
4
for chunk in model.generate_streaming(
5
    text="流式文本转语音非常简单！",
6
):
7
    chunks.append(chunk)
8
wav = np.concatenate(chunks)
9
sf.write("streaming.wav", wav, model.tts_model.sample_rate)

1
from nanovllm_voxcpm import VoxCPM
2
import numpy as np
3
import soundfile as sf
4

5
server = VoxCPM.from_pretrained(model="/path/to/VoxCPM", devices=[0])
6
chunks = list(server.generate(target_text="你好，VoxCPM！"))
7
sf.write("out.wav", np.concatenate(chunks), 48000)
8
server.stop()

1
{"audio": "/path/to/audio_001.wav", "text": "这是第一段训练文本。"}
2
{"audio": "/path/to/audio_002.wav", "text": "这是第二段训练文本。"}

1
import json
2
from voxcpm import VoxCPM
3
from voxcpm.model.voxcpm import LoRAConfig
4
import soundfile as sf
5

6
base_model_path = "openbmb/VoxCPM2"
7
lora_ckpt_dir = "/path/to/checkpoints/lora/step_0002000"
8

9
# 加载 LoRA 配置
10
with open(f"{lora_ckpt_dir}/lora_config.json") as f:
11
    lora_info = json.load(f)
12
lora_info["base_model"] = base_model_path
13
lora_cfg = LoRAConfig(**lora_info["lora_config"])
14

15
# 加载基础模型 + LoRA 适配器
16
model = VoxCPM.from_pretrained(
17
    hf_model_id=base_model_path,
18
    lora_config=lora_cfg,
19
    lora_weights_path=lora_ckpt_dir,
20
)
21

22
# 正常推理
23
wav = model.generate(
24
    text="这是我微调后的声音。",
25
    cfg_value=2.0,
26
    inference_timesteps=10,
27
)
28
sf.write("finetuned_output.wav", wav, model.tts_model.sample_rate)