Speech Generation

Overview

Bark is the text-to-audio model Suno released as open source in April 2023. It runs in three transformer stages: text becomes semantic tokens, those become coarse EnCodec codes, and a third stage fills in the fine codes behind the waveform. There is no phoneme step, so Bark can laugh, sigh or sing a line along with the words. Being fully generative, it can also stray from the script.

On this page Bark reads a block of text in one voice. The text is split into sentences, each is generated with your preset, and the pieces are joined with a short pause, so input is not limited to the 13 to 14 seconds of one Bark pass. The large checkpoint is loaded.

What you can build

• Voiceover drafts for explainers and product demos
• Narration in Hindi, Japanese, Polish, Turkish or another of the 13 languages
• Character lines with a written-in laugh, gasp or hesitation
• Short sung lines, with the lyrics wrapped in ♪
• Read-aloud features for accessibility products

Inputs

• Prompt: the text to read. Bark works out the language from the text. The page lists the cues it responds to: [laughter], [laughs], [sighs], [music], [gasps], [clears throat], "..." for a hesitation and ♪ around lyrics. Suno also lists CAPITALS for stress and [MAN] or [WOMAN] to bias the speaker.
• Speakers: one preset out of 130, ten each for English, Chinese, French, German, Hindi, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish and Turkish, each labelled male or female. English Speaker 0 (male) is the default.
• Advanced settings, under "Show advanced configurations", each from 0 to 1:
  • silence (0.25): the pause between sentences
  • text-temp (0.5): temperature of the text-to-semantic stage
  • waveform-temp (0.7): temperature of the coarse audio-code stage
  • min-eos-p (0.15): how sure Bark must be that a sentence has ended before it stops. Lower values stop sooner and can clip the last word; higher values let it run on

Outputs

One mono MP3 with the whole reading in the chosen voice. One sample on this page runs about 83 seconds, far longer than a single Bark pass.

Recommended settings

• Pick a preset from the same language as the text. When a text mixes languages, Bark tries to give each one its native accent.
• Break up any sentence longer than about 13 seconds of speech; it will not fit in one pass.
• If Bark tacks extra audio onto the end of a sentence, lower min-eos-p a little or generate again, since every run is sampled fresh.
• A lower text-temp gives a steadier read; higher values give more varied takes.

Limitations

• Bark may drop words or add ones that are not in the text; Suno says it varies more than conventional text-to-speech.
• Quality ranges from clean speech to something closer to a bad microphone in a crowded room.
• English sounds best for now.
• Voices come only from the presets. Bark has no custom voice cloning.

Safety & compliance

Bark has been MIT-licensed since Suno's May 2023 update, which allowed commercial use. The model card says it is meant for research and that its output is not censored. Suno takes no responsibility for what Bark generates.