Bark: audio generator — what it does and what you need to run it

Bark is an open-source audio generator from Suno that solves the text-to-speech task: it turns text into speech and other sounds based on a text description. The model and code are publicly available, so you can run it yourself and integrate it into your project.

What it does

Bark works as a text-to-speech model. As described by the author, it is a generative audio model controlled by text: it synthesizes speech and can also create other sounds if that follows from the text prompt. The authors do not provide details about supported languages, voices, or the duration of generated audio in the fact sheet.

The model is available in the transformers library. The weights take up 4.2 GB — this is both the largest file and the total size of all files.

What you need to run it

The code is written primarily in Jupyter Notebook; the last change to the repository was on August 19, 2024. The repository was created on April 7, 2023.

The code license is MIT, and the weights license is MIT. This permits use and modification of the model, including in commercial projects, provided the license terms are preserved.

As described by the author, the full version of Bark requires about 12 GB of VRAM to hold all components on the GPU at the same time. Verbatim: «The full version of Bark requires around 12GB of VRAM to hold everything on GPU at the same time».

The platform as described by the author is CUDA. Bark was tested on CPU and GPU with pytorch 2.0+, CUDA 11.7, and CUDA 12.0. This is a developer statement, not an independent measurement.

Who it suits

Bark suits those looking for an open-source text-to-speech model with an MIT license and willing to run it on a machine with CUDA and sufficient VRAM. The project is convenient for experiments in Jupyter Notebook and for integration via transformers.

The model is unlikely to suit those who do not have a GPU with 12 GB of VRAM or who expect a ready-made cloud service without running it themselves. It is also worth noting that the authors do not provide details about supported languages or audio length limits.

Bark is an open generative audio model that can be run locally if you have a suitable GPU. It is suitable for research tasks and projects where text-to-speech generation with the ability to customize the code is needed.