[Windows] No triton, no worries. Use flash attention 2.

[bukit-kronik]

from voxcpm import VoxCPM
import soundfile as sf
import numpy as np
import librosa

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

segments = [
    "One sentence",
    "or more",
]

all_wavs = []
for seg in segments:
    wav = model.generate(
        text=seg, # add voice control "(speaking like whispering, slow and mysterious tone)" embedded to the each sentence/paragraph.
        prompt_wav_path="path/to/10-30s/cloning.wav", # disable this if you want voice control
        prompt_text="the cloning.wav transcript, punctuation matters.",  # disable this if you want voice control
        reference_wav_path="path/to/10-30s/cloning.wav",  # same wav as prompt_wav_path  
        cfg_value=1.5, #1.5 - 2.0, 2.5 for voice control
        inference_timesteps=10, # 6 fast, acceptable quality   
        normalize=True, # try False
        retry_badcase=True # try False
    )
    wav_trimmed, _ = librosa.effects.trim(wav)
    all_wavs.append(wav_trimmed)

full_wav = np.concatenate(all_wavs)
sf.write("hifi_clone.wav", full_wav, model.tts_model.sample_rate)

print("saved: hifi_clone.wav")

# uv pip install -r requirements.txt

torch==2.8.0+cu129
torchaudio==2.8.0+cu129
voxcpm==2.0.2
soundfile==0.13.1

uv pip install https://huggingface.co/ussoewwin/Flash-Attention-2_for_Windows/blob/main/flash_attn-2.8.2%2Bcu129torch2.8.0cxx11abiTRUE-cp312-cp312-win_amd64.whl

You will get some minor errors, handoff to your favorite coding ai, let it modify the source.

IMO voice control is better than providing transcript, not 100% cloning but the likeness is great.

Thanks to OpenBMB for releasing this open source, hope you can lower the memory consumption and provide faster generation like Kyutai Pocket TTS (CPP Version!).