VoxCPM1.5

VoxCPM Speech Series

VoxCPM1.5 generates Chinese and English speech with 44.1kHz output, prompt-continuation voice cloning and streaming generation. As the upgrade within the 1.x line it keeps the existing usage paths and supports LoRA and full fine-tuning, making it a fit for continuing existing Chinese/English voice projects, customizing a speaker’s voice, or evaluating higher sample-rate output.

Get the model weights and code for VoxCPM1.5. Use the weights and configuration for this version, and review the license and usage notes on the model page before downloading.

VoxCPM1.5 is the final upgrade of the 1.x series, targeting Chinese and English text-to-speech and zero-shot voice cloning by continuing from a prompt audio. It keeps the 1.x usage path while raising the output sample rate to 44.1kHz and lowering the LM token rate from 12.5Hz to 6.25Hz.

Model Downloads

Model	Hugging Face	ModelScope	License
VoxCPM1.5	openbmb/VoxCPM1.5	OpenBMB/VoxCPM1.5	Apache-2.0

Version Information

Item	Official specification
Parameters	800M
Supported languages	Chinese, English
Output sample rate	44.1kHz
LM token rate	6.25Hz
Maximum sequence length	4096 tokens
Cloning method	Prompt-audio continuation: requires a prompt audio and its accurate transcript.
Fine-tuning	Supports SFT and LoRA.
RTX 4090 RTF	About 0.15 per the official version history; actual values depend on the text, parameters, and runtime environment.

When to Use It

Need	Recommendation
You already have VoxCPM 1.x inference or training assets	Prefer VoxCPM1.5 to keep the architecture and interface path continuous.
Chinese-English bilingual with 44.1kHz output	VoxCPM1.5 is the better fit within 1.x.
Multilingual, voice design, isolated-reference cloning, or Style Control	Go straight to VoxCPM2; these are not VoxCPM1.5 features.

Usage Boundaries

VoxCPM1.5 does not have VoxCPM2's reference_wav_path isolated reference channel, and it does not support Voice Design or Style Control. For cloning, pass prompt_wav_path and prompt_text as a pair; a mismatch between the reference text and audio directly affects continuation-based cloning quality.

Quick Start

Install the dependencies and explicitly load VoxCPM1.5. First generate a clip of Chinese/English speech, then try prompt-audio-continuation cloning and streaming output. The examples cover both Python and the command line, and explain how to verify the audio sample rate and reference text.

Installation

Bash

pip install voxcpm soundfile

The current voxcpm package defaults to VoxCPM2. When using VoxCPM1.5, explicitly specify openbmb/VoxCPM1.5 in both Python and the CLI.

Basic Text-to-Speech

Python

from voxcpm import VoxCPM
import soundfile as sf

model = VoxCPM.from_pretrained(
    "openbmb/VoxCPM1.5",
    load_denoiser=False,
)

wav = model.generate(
    text="VoxCPM1.5 支持中文和英文语音合成。",
    cfg_value=2.0,
    inference_timesteps=10,
    seed=42,
)
sf.write("voxcpm15.wav", wav, model.tts_model.sample_rate)

VoxCPM1.5 outputs at a 44.1kHz sample rate. Use the sample_rate from the model object when writing files, to avoid getting the sample rate wrong by hand.

Voice Cloning

VoxCPM1.5 clones voices through prompt audio continuation. prompt_wav_path and prompt_text must be provided as a pair, and the transcription should match the audio word for word.

Python

wav = model.generate(
    text="这是使用 VoxCPM1.5 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("voxcpm15_clone.wav", wav, model.tts_model.sample_rate)

Streaming Output

Python

import numpy as np

chunks = []
for chunk in model.generate_streaming(
    text="这是 VoxCPM1.5 的流式输出。",
):
    chunks.append(chunk)

wav = np.concatenate(chunks)
sf.write("voxcpm15_stream.wav", wav, model.tts_model.sample_rate)

Command Line

Bash

# 基础合成
voxcpm design   --hf-model-id openbmb/VoxCPM1.5   --text "VoxCPM1.5 supports Chinese and English."   --output voxcpm15.wav

# 提示音频续写式克隆
voxcpm clone   --hf-model-id openbmb/VoxCPM1.5   --text "这是克隆后的目标文本。"   --prompt-audio speaker.wav   --prompt-text "提示音频中实际说出的完整文本。"   --output voxcpm15_clone.wav

Here the CLI subcommand design serves only as the entry point for basic synthesis; it does not mean VoxCPM1.5 has VoxCPM2's voice design capability. Don't use --reference-audio with 1.5; the current CLI rejects that combination.

Parameter Tuning

Parameter	Purpose	Troubleshooting
cfg_value	Controls guidance strength between the text and diffusion generation	Lower it if the voice sounds tense or distorted; raise it slightly if short sentences are unclear.
inference_timesteps	Controls the number of diffusion solver steps	Increasing it raises compute cost; use the default of 10 as a baseline first.
normalize	Enables external text normalization	If numbers or abbreviations are handled incorrectly, compare results with it on and off.
denoise	Reference audio enhancement	The model card notes it may introduce distortion and limits the sample rate to 16kHz; for high-fidelity cloning, compare against results with it off.

FAQ

Problem	Solution
Mistakenly used reference_wav_path	Use prompt_wav_path and prompt_text instead.
Low cloning similarity	Make sure the audio is single-speaker and low-noise and the prompt text matches word for word; compare results with and without denoising.
Output isn't 44.1kHz	Confirm you loaded VoxCPM1.5, and save with model.tts_model.sample_rate.
Unstable results in other languages	VoxCPM1.5 officially supports only Chinese and English; other languages are not guaranteed.

Related Resources

Topic	Link
Deployment	VoxCPM1.5 deployment guide
Fine-tuning	VoxCPM1.5 fine-tuning guide
Technical report	VoxCPM Technical Report
Code repository	OpenBMB/VoxCPM