bark

What does Bark generate? Speech, music, background noise, simple sound effects, and even laughter, sighs, and crying, all from text.

Bark is a transformer based text to audio model created by Suno. The headline ability is highly realistic multilingual speech, but the output list goes further: music, background noise, and simple sound effects. It also handles nonverbal sounds, like laughing, sighing, and crying. Because Bark is fully generative, it does not behave like a conventional text to speech engine, and the README says output can deviate from prompts in unexpected ways. A disclaimer frames the whole thing as research software, with Suno taking no responsibility for generated output.

Pretrained checkpoints are provided for the research community, ready for inference and usable commercially. A May 2023 update moved Bark to the MIT License, doubling GPU speed and multiplying CPU inference by ten. A smaller model variant trades a little quality for more speed. Later additions cover long form generation, voice consistency improvements, and a voice prompt library. Users hunting for text to music models get pointed to Suno's web page instead, and a Discord community exists for sharing prompts and presets.

Bark supports a range of languages out of the box and guesses the language from the input text. With code-switched text it tries to use the native accent of each language, and English quality is described as the best for now. The README warns that a German history prompt typed in English will typically produce English audio with a German accent. Music is treated like any other audio, and sometimes Bark decides to generate text as music, so adding music notes around lyrics can help steer it.

Bark is positioned as research software with a permissive license and a long feature list. The caveats about output variance are printed right in the README, and they are worth reading before building anything on top of it.

• Source 1 · Project repository
• Source 2 · README