Bark

Bark is Suno's text-to-audio model for developers and researchers experimenting with speech and other sounds. It can depart from the supplied script, which matters if you need an exact spoken reading.

• Multilingual speech. Generate speech in English, German, Spanish, French, Hindi, Italian, Japanese, Korean, Polish, Portuguese, Russian, Turkish and simplified Chinese, with automatic language detection.
• Expressive sounds. Include laughter, sighs, crying and other nonverbal sounds alongside speech, or generate music, background noise and simple sound effects.
• Mixed-language text. Provide text in more than one language and the model attempts to use the corresponding native accents.
• Longer audio. Default generation works well with about 13 seconds of spoken text; notebook examples cover longer output.

Bark is open source under MIT, and its pretrained checkpoints are available for commercial use. It runs on your hardware through Python or Hugging Face Transformers and supports CPU and CUDA GPUs. CPU and older GPU generation can be significantly slower.

The full model needs about 12 GB of VRAM to keep everything on the GPU. Smaller models fit in about 8 GB. Bark was developed for research and demonstrations, and its generative output can vary unexpectedly from the prompt.