AI Vocals on lyria-3.pro
The workbench can generate a track with sung vocals from a text prompt or from lyrics you supply. It is not a voice-cloning product, and it does not promise a specific singer’s timbre. The vocal you get depends on the selected model, the language, and the direction you write.
Two ways to start
Prompt-only vocals. Describe the voice and the song: “airy female lead, close-mic indie folk, 92 BPM, restrained chorus.” The model drafts both the words and the performance.
Lyrics-to-vocals. Paste your lyrics with a Lyrics: prefix, then add style and vocal direction. This is the better path when the words are the point — a birthday song, a proposal, or a lyric-first rap draft. The lyrics workflow covers line shape and syllable counts.
What you can direct
- Range and timbre — airy soprano, textured alto, warm baritone.
- Delivery — whispered, belted, talk-singing, rhythmic chant.
- Language — name it explicitly; multilingual output is model-dependent.
- Arrangement — where the vocal enters, whether there are backing harmonies, and how the last chorus resolves.
You cannot direct a specific real person’s voice. Do not upload a reference track and ask for an imitation of a named artist; that is a rights problem, not a prompt trick.
The vocal workflow
- Decide whether the words or the mood come first. Write the lyric first if the song is for a person or an event. Write the mood first if you are scoring a scene.
- Pick a model and read the cost. The picker shows the provider and credit price. Suno lanes cost 5–12 credits per draft; MiniMax Music 2.5 costs 30. The Google Lyria lane is disabled on this deployment and is not sold.
- Generate one take and check intelligibility. Can you hear the consonants? Is the name pronounced correctly? Fix the lyric or the vocal direction before adding layers.
- Use the editing tools only after the take works. Extend, add tracks, and vocal removal are separate billed tasks. Read the AI Music Editor page before stacking edits.
Rights and disclosure
Generated vocals are not a licence to imitate a real artist. Commercial use depends on the provider terms, your plan, your inputs, and applicable law. If you publish a song with AI vocals, disclose it where the platform or the client requires disclosure. The Terms of Service state the platform’s boundary; they do not grant rights the provider does not grant.
FAQ
Can I clone my own voice?
The workbench has a separate voice-cloning entry with its own requirements and cost. Only upload a sample you have the right to use. This page is about generated vocals, not cloning.
Can I get an instrumental version of a vocal track?
Yes. The Vocal Remover splits a generated track into vocal and instrumental parts. Separation quality depends on the source mix.
Which languages are supported?
It depends on the selected model. Name the language in the prompt and test a short take before writing a full song. Do not assume every lane supports every language.
Why do the words sound wrong?
Models can mispronounce names, drop syllables, or rush a line. Shorten the line, spell the name phonetically in the lyric, and regenerate. Keep the take that pronounces the important words correctly.