OpenBMB has introduced the VoxCPM model for speech generation

The OpenBMB team has released VoxCPM2, a specialized text-to-speech (TTS) model. The system is built on an architecture with 2 billion parameters and operates without a tokenizer, allowing it to process audio data directly.

The model supports audio generation in 30 languages and provides audio output with a 48 kHz sampling rate. One of the key features of the release is the Voice Design function, which allows users to flexibly customize the characteristics of the synthesized voice for specific tasks.

The development is focused on creating high-quality audio content with minimal latency. The absence of an intermediate tokenization stage simplifies the data processing pipeline, making the model suitable for integration into interactive systems that require natural sound and multilingual support.

For developers and researchers, access to the code and model weights is open in the project's official repository. The tool allows for the implementation of advanced speech synthesis capabilities in local applications, expanding the toolkit for working with audio in tasks that require high precision in intonation and support for a wide range of language pairs.