voxcpm

What it is and what it does

VoxCPM2 is a large language model–based text-to-speech system that bypasses traditional discrete tokenization, instead generating continuous speech representations end-to-end. It outputs 48kHz studio-quality audio and supports 30 languages including Chinese dialects, with no language tags required. The model is built on a MiniCPM-4 backbone and trained on over 2 million hours of multilingual speech data.

The package offers three primary synthesis modes: voice design (create a voice from natural-language description alone), controllable voice cloning (clone a voice from a short reference clip with optional style guidance), and ultimate cloning (provide both reference audio and transcript for seamless continuation with full vocal nuance preservation). It also supports streaming generation for real-time applications. All synthesis is context-aware, automatically inferring appropriate prosody and expressiveness from text content.

Use it for:

• Generate natural multilingual speech for applications serving users in 30 languages without language-specific models or tags.
• Create synthetic voices from text descriptions for brand narration, character voices, or accessibility without reference audio.
• Clone a speaker's voice from a short clip and adjust emotion, pace, or tone while preserving original timbre.
• Build real-time speech synthesis pipelines using streaming generation with low latency on modern GPUs.
• Reproduce exact vocal characteristics (timbre, rhythm, emotion) by providing reference audio and its transcript for high-fidelity cloning.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

VoxCPM2 is a tokenizer-free text-to-speech system that generates multilingual speech directly via diffusion-autoregressive architecture, supporting voice design, controllable voice cloning, and 48kHz audio output across 30 languages.

Yes, with conditions. VoxCPM2 is worth installing if you need multilingual TTS with voice cloning and design capabilities, have access to a modern GPU (CUDA ≥12.0), and can accommodate 23 dependencies including torch and transformers. The Apache-2.0 license permits commercial use. Active maintenance and no known vulnerabilities are positive signals. The main friction is the large dependency footprint and GPU requirement; it is not suitable for CPU-only or lightweight environments.

Install

voxcpm on PyPI

pip install voxcpm

uv add voxcpm

poetry add voxcpm

Installing voxcpm

Before you install

Low install friction with a pure-Python wheel, but carries 23 runtime dependencies including torch, transformers, and audio libraries. Active maintenance with recent releases; requires Python ≥3.10 and PyTorch ≥2.5.0.

License in practice

Apache-2.0 permissive license allows free commercial use without restriction, making the package suitable for both open-source and proprietary applications.

Quickstart

pip install voxcpm

from voxcpm import VoxCPM
import soundfile as sf

model = VoxCPM.from_pretrained("openbmb/VoxCPM2", load_denoiser=False)
wav = model.generate(text="Hello world", cfg_value=2.0, inference_timesteps=10)
sf.write("output.wav", wav, model.tts_model.sample_rate)

Requires Python ≥3.10 (<3.13), PyTorch ≥2.5.0, and CUDA ≥12.0; model weights must be downloaded from HuggingFace or ModelScope.

Verify before relying

• Actual inference speed (RTF) on consumer hardware; documentation cites RTF ~0.3 on RTX 4090 but real-world performance varies.
• Memory requirements for loading the 2B parameter model on different GPU tiers.
• Quality and naturalness of voice design outputs compared to reference-based cloning.
• Multilingual synthesis quality parity across all 30 supported languages.

Package facts

Evidence: voxcpm-2.0.3-py3-none-any.whl

Keywords: voxcpm, text-to-speech, tts, speech-synthesis, voice-cloning, ai, deep-learning, pytorch

Tags

More Artificial Intelligence packages

Further reading