VoxCPM2 for Multilingual Product Narration: Hear Aurora in Three Languages
At OpenBMB, we built VoxCPM2 to generate speech in 30 supported languages. The 2B-parameter model produces native 48 kHz audio and supports both voice design from a written description and reference-based voice cloning.
Below, the same example product, the fictional Aurora Desk Light, is introduced in English, Chinese, and Spanish. Each script is included in full, so you can follow the product name, button instructions, and ending as you read.
Hear the Aurora product narration in three languages
English — 24.00 seconds
Meet the Aurora Desk Light, a simple way to bring focused light to your workspace. Start your morning by opening your notebook, setting out your plans, and tapping the button once. A focused beam puts the page in front of you at the center of attention. When you need a different level of brightness, hold the button and adjust the light to suit the task. Use this moment to settle in, review your priorities, and choose where to begin. At the end of the day, give your lamp a full charge and get ready for your next work session.
Take selection mattered here: an earlier short English version had audible noise and sharp syllables. We replaced it with this longer take after listening.
Chinese — 25.60 seconds
认识一下 Aurora 台灯，用一束聚焦的光照亮你的工作空间。早晨开始时，打开笔记本，整理好当天的计划，再轻按一次按钮。聚焦光束会照亮你面前的页面。需要改变亮度时，长按按钮，让光线适合手头的任务。利用这片刻时间安顿下来，回顾待办事项，并决定先从哪里开始。一天结束后，为台灯充满电，准备迎接下一次工作。
Listening to this Chinese take revealed no obvious problems with the product name, skipped or repeated words, pauses, or the ending.
Spanish — 29.76 seconds
Conoce la lámpara de escritorio Aurora, una forma sencilla de iluminar tu espacio de trabajo. Empieza la mañana abriendo tu cuaderno, organizando tus planes y pulsando el botón una vez. Un haz de luz concentrado ilumina la página que tienes delante. Cuando necesites otro nivel de brillo, mantén pulsado el botón y ajusta la luz a tu tarea. Aprovecha ese momento para acomodarte, repasar tus prioridades y decidir por dónde empezar. Al terminar el día, carga por completo la lámpara y prepárate para tu próxima sesión de trabajo.
The listening check found no obvious noise, repeated passages, or sudden ending in this take.
These three samples use plain text-to-speech: no voice description and no reference recording. They demonstrate the same product narrative in three languages; they do not use a shared cloned speaker.
When to use Voice Design or a reference voice
Voice Design adds a written description before the spoken text. For example, the documented syntax places a description in parentheses:
text = "(A young woman, gentle and sweet voice)Meet the Aurora Desk Light."
This is an input example, separate from the three plain TTS recordings above. A description requests a delivery; the generated take still determines whether it suits the product. Our model card recommends generating one to three candidates because results can vary.
If you already have an established spokesperson, reference-based cloning is the relevant workflow. Use a recording with permission covering the intended use. The official demo page provides Voice Design and cloning examples to hear alongside the local product samples here.
What the published benchmarks add
Our Seed-TTS-eval results provide context for speech accuracy and reference-speaker similarity:
Evaluation
Transcription error
Speaker similarity
English
1.84% word error rate
75.3%
Chinese
0.97% character error rate
79.5%
These figures belong to the published benchmark, rather than a scoring exercise on the Aurora recordings. For this product example, the audio and paired scripts are the direct evidence to inspect.
The deployment table lists about 8 GB of GPU memory and a real-time factor of about 0.30 on an RTX 4090; the Nano-vLLM path reports about 0.13. A factor below 1 means the audio renders faster than its playback duration in that setup. Those GPU figures describe a different environment from our local CPU generation.
Try your own product script
Install the Python package and save a generated waveform using the model’s own sample rate:
pip install "voxcpm==2.0.3" soundfile
import soundfile as sf
from voxcpm import VoxCPM
model = VoxCPM.from_pretrained(
"openbmb/VoxCPM2",
device="cpu", # Use "cuda" for a compatible NVIDIA GPU.
load_denoiser=False,
optimize=False,
)
text = "Meet the Aurora Desk Light. Tap once for a focused beam."
wav = model.generate(
text=text,
cfg_value=2.0,
inference_timesteps=10,
max_len=384,
normalize=False,
denoise=False,
)
sf.write("aurora.wav", wav, model.tts_model.sample_rate)
Replace text with a script in the language you need. For Voice Design, add the parenthesized description shown above. The example loads the model once; reuse that instance for additional scripts. The first run downloads the model weights.
Our longer Aurora files were generated locally with voxcpm 2.0.3 on CPU in bfloat16, using seed 42, 10 inference steps, guidance 2.0, and a generation limit of 384. No denoiser or reference audio was used. The selected English file took about 305 seconds to generate 24 seconds of audio, excluding model loading; this was an offline rendering workflow.
From a sample to your own narration
Use the Aurora scripts as a starting point, then substitute your own product name and instructions. Listen through the ending as well as the opening, and retain the script and settings with the take you choose. For longer material, a sequence of approved sections is easier to revise than one uninterrupted recording.
If you also evaluate alternatives such as Qwen3-TTS or IndexTTS, use the same product scripts. The recordings here demonstrate VoxCPM2; they do not establish a ranking between systems.
Appendix: the earlier one-sentence sample
Earlier sample: “OpenBMB builds open-source AI models for everyone.” This 3.52-second, 48 kHz recording used plain text-to-speech with no voice description or reference recording. It is retained here for reference; the longer three-language Aurora recordings are the main product demonstration.
Sources
Features, usage syntax, benchmark results, and deployment figures were checked on 22 October 2026 against the VoxCPM repository, VoxCPM2 model card, and official audio demo. The three Aurora scripts and the generation settings are included above for reference.