VoxCPM2

VoxCPM2 Frequently Asked Questions

What is VoxCPM2 and how does it differ from other TTS models?

VoxCPM2 is an open-source tokenizer-free text-to-speech model with 2B parameters developed by OpenBMB. Unlike traditional TTS systems, VoxCPM2 uses diffusion autoregressive architecture to generate continuous speech representations directly. VoxCPM2 uniquely combines zero-shot voice design from text descriptions with controllable voice cloning, offering three distinct modes: voice design without reference audio, controllable cloning with style control, and ultimate cloning with transcript guidance for maximum fidelity.

What languages does VoxCPM2 support?

VoxCPM2 supports 30 languages: Arabic, Burmese, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Khmer, Korean, Lao, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Tagalog, Thai, Turkish, and Vietnamese. VoxCPM2 also includes Chinese dialect support for Cantonese, Sichuanese, Wu, Northeastern, Henan, Shaanxi, Shandong, Tianjin, and Minnan dialects. VoxCPM2 automatically detects input language in most cases.

How does VoxCPM2 voice design work without reference audio?

VoxCPM2's voice design mode uses zero-shot learning to create entirely new voices from natural language descriptions. Simply describe desired voice characteristics in parentheses at the start of your text—such as gender, age, accent, pitch, speaking style—and VoxCPM2 synthesizes a matching voice instantly. This eliminates the need for voice talent recording or reference audio collection, making custom voice creation accessible for character development, brand voices, and creative projects.

What audio quality does VoxCPM2 produce?

VoxCPM2 outputs studio-quality 48kHz audio suitable for professional media production. VoxCPM2 uses AudioVAE V2 with asymmetric encode/decode architecture that accepts 16kHz reference audio and outputs 48kHz with built-in super-resolution. This high-fidelity output eliminates the need for external upsampling or post-processing. VoxCPM2 achieves state-of-the-art results on major TTS benchmarks including Seed-TTS-eval, CV3-eval, and InstructTTSEval.

What are VoxCPM2's hardware requirements?

VoxCPM2 requires approximately 8GB VRAM for inference with the 2B parameter model in bfloat16 precision. VoxCPM2 achieves RTF of 0.30 on RTX 4090 GPU, or 0.13 with Nano-VLLM optimization for faster generation. Minimum requirements are Python 3.10+, PyTorch 2.5.0+, and CUDA 12.0+. VoxCPM2 can run on consumer-grade GPUs, making professional voice synthesis accessible to individual developers and small teams without enterprise infrastructure.

Is VoxCPM2 free to use for commercial projects?

Yes, VoxCPM2 is fully open source under Apache 2.0 license, allowing both personal and commercial use. You can download VoxCPM2 model weights from Hugging Face, deploy on your own infrastructure, and customize for your specific needs. VoxCPM2 supports fine-tuning via supervised fine-tuning (SFT) and LoRA for voice adaptation. OpenBMB provides complete documentation, code, and model weights with no API fees or usage restrictions for VoxCPM2.

Can I control emotion and speaking style with VoxCPM2?

Yes, VoxCPM2's controllable cloning mode provides fine-grained control over voice attributes. You can adjust emotion, pace, pitch variation, and speaking style while preserving the original voice timbre from reference audio. VoxCPM2 accepts natural language control instructions to steer voice characteristics. Note that controllable generation results may vary between runs—VoxCPM2 developers recommend generating 1-3 times to achieve desired voice or style as they continue improving controllability consistency.