OpenBMB's VoxCPM2, officially released in April 2026, is an innovative next-generation text-to-speech system based on open-source technology. It generates natural-sounding speech within a continuous acoustic space, as if a person were speaking directly. To overcome the limitations of existing speech models, which divided subtle sound units into mechanical tokens, leading to computational bottlenecks or unnatural word combinations, it organically connects the MiniCPM large language model backbone with an improved high-resolution audio encoder architecture. By departing from the conventional discontinuous audio token transformation process, which converts text into mechanical sound fragments and outputs them, it introduces a tokenizer-free architecture that directly reconstructs the continuous waveform and texture of the sound, enabling the realization of 48kHz studio-quality audio output even in local environments. While traditional text-to-speech methods involve breaking down text words into small units, converting them into mechanical pronunciation symbols or discontinuous tokens, and then synthesizing them into audio, VoxCPM2 directly infers a continuous acoustic space, creating a natural waveform with no breaks or artificial tonal distortions. Just as a large language model reads the flow of previous context and naturally continues a sentence, VoxCPM2 organically weaves together the subtle breaths and accents from the beginning to the end of the sound. It offers voice design, which allows users to create new voice personas simply by describing emotions and vocal characteristics in text, and zero-shot voice cloning, which replicates a unique timbre with just a short audio clip. This model provides both powerful performance and flexibility for researchers building global services that require multilingual support or artificial intelligence voice agents. Researchers can use VoxCPM2 to transmit synthesized speech that naturally transcends languages in real-time streaming from over 30 major multilingual datasets. It can be widely used, from leveraging local GPU resources to mass-produce emotionally rich audio from text scripts in just a few seconds, to developing customized pipelines that utilize efficient fine-tuning techniques such as LoRA to precisely learn even the subtle linguistic habits of a specific speaker.

Open-source, next-generation text-to-speech model that generates natural-sounding speech in a continuous acoustic space.