VoxCPM-0.5B

VoxCPM Speech Series

Following the runtime requirements of the first-generation model, choose Python integration, CLI batch processing or offline deployment. Pin the model and runtime code versions, confirm the 16kHz output, and set up quality and performance regression checks for your existing project.

VoxCPM-0.5B is the first-generation 1.0 checkpoint, suited to reproducing historical baselines, maintaining legacy projects, or validating 16kHz Chinese and English speech. For new projects without compatibility constraints, evaluate VoxCPM2 first; if you need a higher sample rate but want to stay on the 1.x path, evaluate VoxCPM1.5.

Deployment information

Item	Value or notes
Model	openbmb/VoxCPM-0.5B
Languages	Chinese, English
Output	16kHz
Reference VRAM	About 5 GB per the official series README; actual usage depends on runtime parameters.
RTX 4090 RTF	About 0.17 per the official version history.
Cloning	Prompt audio plus an accurate transcript; isolated reference audio is not supported.

Python integration

Python

from voxcpm import VoxCPM
import soundfile as sf

model = VoxCPM.from_pretrained(
    "openbmb/VoxCPM-0.5B",
    load_denoiser=False,
)

def synthesize(text: str, output: str) -> str:
    wav = model.generate(
        text=text,
        cfg_value=2.0,
        inference_timesteps=10,
    )
    sf.write(output, wav, model.tts_model.sample_rate)
    return output

Load the model only once, at process startup. In batch or service scenarios, the application layer handles concurrency, queuing, timeouts, retries and file cleanup.

CLI and offline operation

Bash

# 在线加载
voxcpm design   --hf-model-id openbmb/VoxCPM-0.5B   --text "部署验证。"   --output out.wav

# 批处理
voxcpm batch   --hf-model-id openbmb/VoxCPM-0.5B   --input texts.txt   --output-dir outputs

Python

from huggingface_hub import snapshot_download

snapshot_download(
    "openbmb/VoxCPM-0.5B",
    local_dir="./models/VoxCPM-0.5B",
)

Bash

voxcpm design   --model-path ./models/VoxCPM-0.5B   --local-files-only   --text "离线验证。"   --output offline.wav

Batch cloning

Bash

voxcpm batch   --hf-model-id openbmb/VoxCPM-0.5B   --input texts.txt   --output-dir clone_outputs   --prompt-audio speaker.wav   --prompt-text "提示音频中实际说出的完整文本。"

Do not use --reference-audio; that parameter applies only to VoxCPM2. Batch outputs should be verified file by file for duration, empty audio, abnormal noise and text intelligibility.

Ecosystem backend boundaries

The official series README lists a reference RTF of about 0.10 for 0.5B on an RTX 4090 with Nano-vLLM. That backend is maintained independently by the community, so you must check its repository for current 0.5B support before using it. The current official examples for vLLM-Omni and llama.cpp-omni focus on VoxCPM2, so you cannot assume from them that 0.5B runs directly.

Production checks

Dimension	Check
Versioning	Pin the model revision, code commit, dependency lock file and image digest.
Quality	Maintain a fixed regression set covering Chinese and English, numbers, long and short sentences and cloning; listen and run automated checks before and after upgrades.
Performance	Record RTF, first-packet and total latency, concurrency, peak VRAM and failure rate.
Output	Confirm the WAV files play and are 16kHz; resample explicitly when downstream needs a different sample rate.
Safety	Authentication, length limits, speaker authorization, generation labeling and minimal audio logging.