VoxCPM-0.5B

VoxCPM Speech Series

VoxCPM-0.5B is the first public VoxCPM model. It supports Chinese and English text-to-speech, prompt-continuation voice cloning and streaming generation, with 16kHz output. It keeps the original model’s usage and fine-tuning paths, and suits reproducing historical baselines, maintaining existing projects, or validating Chinese/English speech generation.

Get the model weights and runtime code for VoxCPM-0.5B. Use the weights and configuration for this version, and review the license and usage notes on the model page before downloading.

VoxCPM-0.5B is the original public checkpoint of VoxCPM 1.0 and the historical baseline of the tokenizer-free VoxCPM architecture (no discrete audio tokenizer). It targets Chinese and English text-to-speech and prompt-audio-continuation voice cloning, and outputs 16kHz audio.

Model Downloads

Model	Hugging Face	ModelScope	License
VoxCPM-0.5B	openbmb/VoxCPM-0.5B	OpenBMB/VoxCPM-0.5B	Apache-2.0

Version Information

Item	Official Specification
Parameters	640M
Supported languages	Chinese, English
Output sample rate	16kHz
LM token rate	12.5Hz
Maximum sequence length	4096 tokens
Cloning method	Prompt audio continuation: requires prompt audio and its accurate text transcription.
Fine-tuning	Supports SFT and LoRA.
RTX 4090 RTF	Historical reference value for the official release is about 0.17; actual values must be measured on the target machine.

When to Use It

Need	Recommendation
Reproduce the original model's behavior or compare against the early baseline	Use VoxCPM-0.5B.
Maintain existing projects that depend on the original checkpoint	Stay on the current version, and pin the model, Python package and config files to reproducible versions.
Higher-sample-rate Chinese/English output	Evaluate migrating to VoxCPM1.5.
Multilingual support, voice design or controllable voice cloning	Use VoxCPM2.

Limitations

VoxCPM-0.5B does not support VoxCPM2's Voice Design, Style Control or isolated reference-audio cloning. Its cloning interface uses prompt_wav_path and prompt_text; when the audio and transcription don't match, or the audio is too long or very noisy, output stability and similarity will degrade.

Quick Start

Load VoxCPM-0.5B explicitly and complete a first verification of Chinese and English text-to-speech, voice cloning and streaming output. First confirm that the generated audio plays and has the correct sample rate, then follow the guide to check that the reference audio matches its transcript.

Installation

Bash

pip install voxcpm soundfile

Basic text-to-speech

Python

from voxcpm import VoxCPM
import soundfile as sf

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

wav = model.generate(
    text="VoxCPM-0.5B 是初代中英语音生成模型。",
    cfg_value=2.0,
    inference_timesteps=10,
    seed=42,
)
sf.write("voxcpm05.wav", wav, model.tts_model.sample_rate)

VoxCPM-0.5B outputs 16kHz audio. Play the file and check text intelligibility, pauses, volume and abnormal noise.

Voice cloning

Python

wav = model.generate(
    text="这是使用 VoxCPM-0.5B 生成的克隆语音。",
    prompt_wav_path="speaker.wav",
    prompt_text="提示音频中实际说出的完整文本。",
    cfg_value=2.0,
    inference_timesteps=10,
    seed=42,
)
sf.write("voxcpm05_clone.wav", wav, model.tts_model.sample_rate)

The prompt audio and prompt text must be provided as a pair. An inaccurate transcript, multiple speakers, heavy noise or long clips will all reduce the stability of continuation-style cloning.

Streaming output

Python

import numpy as np

chunks = []
for chunk in model.generate_streaming(
    text="这是 VoxCPM-0.5B 的流式输出。",
):
    chunks.append(chunk)

wav = np.concatenate(chunks)
sf.write("voxcpm05_stream.wav", wav, model.tts_model.sample_rate)

Command line

Bash

# 基础合成
voxcpm design   --hf-model-id openbmb/VoxCPM-0.5B   --text "Hello from VoxCPM."   --output voxcpm05.wav

# 提示音频续写式克隆
voxcpm clone   --hf-model-id openbmb/VoxCPM-0.5B   --text "这是目标文本。"   --prompt-audio speaker.wav   --prompt-text "提示音频中实际说出的完整文本。"   --output voxcpm05_clone.wav

The current CLI defaults to VoxCPM2, so you must pass --hf-model-id openbmb/VoxCPM-0.5B explicitly. For 0.5B, the design subcommand simply means basic generation without a reference.

FAQ

Issue	Solution
Used reference_wav_path by mistake	Use prompt_wav_path and prompt_text instead.
Output sample rate is wrong	Confirm the model ID and save using model.tts_model.sample_rate.
Low quality in languages other than Chinese and English	Officially only Chinese and English are supported; use VoxCPM2 for multilingual tasks.
Want to control age, emotion or speaking rate	0.5B has no Voice Design / Style Control; use VoxCPM2 if you need these capabilities.
Cloned voice doesn't sound similar	Switch to a clearer single-speaker reference, proofread the prompt text word by word, and fix the seed for comparison.

Related Resources

Topic	Link
Deployment	VoxCPM-0.5B deployment guide
Fine-tuning	VoxCPM-0.5B fine-tuning guide
Technical report	Technical report in this knowledge base / arXiv paper
Code repository	OpenBMB/VoxCPM