Inworld TTS vs VoxCPM

Inworld TTS has no write-up on Specifiction yet; its figures are above.

VoxCPM is an open-source text-to-speech system whose VoxCPM2 model generates expressive multilingual speech. It supports 30 languages and lists nine Chinese dialects. Voice Design can create a voice from a natural-language description without reference audio, and voice cloning is supported. Output is 48kHz audio, reference audio is accepted at 16kHz, and WAV is listed as an export format. A streaming generation API is available; the repository reports a real-time factor as low as about 0.3 on an NVIDIA RTX 4090. Full and LoRA fine-tuning can adapt the model from 5–10 minutes of audio. Documented deployment includes Python and CLI use, a local web demo, production serving and on-device inference. vLLM-Omni exposes an OpenAI-compatible /v1/audio/speech endpoint, while Nano-vLLM supports concurrent requests. Apple Silicon MPS and C++ inference on CPU, Metal, CUDA or Vulkan are documented. Code and weights use Apache-2.0 and are free for commercial use. Python installation requires Python 3.10 or later but below 3.13, PyTorch 2.5.0 or later and CUDA 12.0 or later. Voice Design and cloning results can vary, and unsupported languages may require testing or fine-tuning. VoxCPM is headquartered in Beijing, China.