VoxCPM2 drops tokenizer-free 48kHz TTS

VoxCPM2's tokenizer-free architecture is a major step forward for natural speech synthesis, bypassing the artifacts often found in discrete token-based systems.

• –Latent-space diffusion autoregressive paradigm enables more expressive and nuanced vocal reproduction than traditional models
• –"Voice Design" creates custom speakers from natural language prompts, a critical tool for creators seeking unique brand identities
• –Native 48kHz output with integrated super-resolution simplifies the production pipeline by removing the need for external vocoders
• –Real-time factor of 0.3 on consumer GPUs makes it viable for streaming and interactive applications
• –Released under Apache-2.0 with open weights, providing a powerful, commercially-friendly alternative to closed-source TTS APIs