VoxCPM2

Why it's ranked here

The project covers unusually broad speech workflows through one model and exposes them through Python, a command line tool, and a browser demo. Its strongest case is practical control: users can steer expression, preserve a reference timbre, or prioritize detailed continuation from transcribed audio.

What's good

It outputs 48 kHz audio directly, accepts 16 kHz reference clips, and includes built-in super-resolution. Generation supports fixed seeds, optional reference denoising, text normalization, streaming chunks, batch processing, and optional word or character timestamps. The same package also includes LoRA training support.

Tradeoffs

The model has two billion parameters, while the documented quick-start stack requires recent Python, PyTorch, and CUDA versions. CPU and Apple MPS are supported in device selection, but optimization only runs on CUDA. MPS defaults to float32 because lower precision can destabilize generation. Transcript-guided cloning cannot be combined with style control.

How to use it well

Use it for local or hosted speech production where multilingual output, designed voices, or controlled cloning matter. Start with the browser demo, then automate repeatable jobs through Python or the command line. Use fixed seeds for reproducibility and transcript-guided continuation when fidelity matters most. It does not provide speech recognition as its primary product.

Observed

License

Apache-2.0 for code and weights

Primary language

Python

Packaging

Setuptools package installable as voxcpm with pip

Interfaces

Python library, command line tool, and Gradio web demo

Runtime devices

CUDA, Apple MPS, and CPU

Python support

Python 3.10 or newer; README limits the quick start to versions below 3.13

Optional surface

Timestamp alignment is available through the timestamps package extra