voxcpm

VoxCPM2 is an open-source text-to-speech system from OpenBMB that avoids discrete tokenization, generating continuous speech representations through an end-to-end diffusion autoregressive architecture. The latest 2B-parameter release is trained on multilingual speech data and supports 30 languages, including Chinese dialects such as Sichuanese, Cantonese, and Wu. Key capabilities described in the README include multilingual synthesis without language tags, voice design from natural-language descriptions (gender, age, tone, emotion, pace), controllable voice cloning from a short reference clip with optional style guidance, and an "ultimate cloning" mode that uses both reference audio and its transcript to reproduce vocal nuances. AudioVAE V2's asymmetric encode/decode design accepts 16kHz reference audio and outputs 48kHz audio with built-in super-resolution. The project provides a Python API, a CLI, and a web demo, plus streaming generation. Deployment options include Nano-vLLM for high-throughput serving, vLLM-Omni for an OpenAI-compatible /v1/audio/speech endpoint, and llama.cpp-omni for on-device C++ inference with GGUF weights on CPU, Metal, CUDA, or Vulkan. Fine-tuning via SFT and LoRA is supported. Weights and code are released under Apache-2.0 for commercial use. The README also lists benchmark results on Seed-TTS-eval and notes risks and limitations. This overview is based solely on the repository's README and does not independently verify performance claims.