google/lyria-3.5

Generate full-length songs with vocals and lyrics from text prompts or images with Lyria 3.5, Google's most capable music generation model

33 runs
Playground API Examples README

Lyria 3.5

Lyria 3.5 is Google DeepMind’s music generation model. It writes and performs full songs from a text prompt and up to 10 images. Songs can have vocals and lyrics, and sections such as verses, choruses, and bridges. Google’s model card calls it “Google’s most capable music generation model to date.”

Lyria 3.5 is the successor to Lyria 3 Pro, which Google’s pricing page now lists as a legacy model. For 30-second clips, use Lyria 3.

What’s new

Google lists these improvements in Lyria 3.5:

  • Richer, more complex melodies that sound more natural
  • Better lyrics that follow your prompt and the song structure more closely
  • More expressive, emotionally nuanced vocals with better pronunciation
  • Easier control over tempo and song length

Inputs

  • prompt describes the song. It can include your own lyrics, section tags like [Verse] and [Chorus], and timestamps like [0:00 - 0:30].
  • images takes up to 10 images for the song to draw on.
  • seed makes results more repeatable. The same seed does not guarantee the same song.
  • output_format is mp3 (the default) or wav.

Output

Each prediction returns one audio file:

  • MP3 by default, at 44.1 kHz stereo. Our test MP3 was 192 kbps.
  • WAV if you set output_format to wav, at 16-bit, 48 kHz stereo.

Google says songs run up to about 3 minutes. With no length in the prompt, our test songs ran about 2 minutes 50 seconds. In our tests, a song took 35 to 50 seconds to generate.

Lyria 3.5 also writes out the lyrics and song structure it used. You can read them in the prediction logs. They are not part of the output.

Write a prompt

A short phrase works, and so does a detailed brief. This one-line prompt comes from Google’s prompt guide:

A folk song about cute cats avoiding puddles, female vocals, acoustic guitar, sound of rain

Google’s advice is to be specific, because vague prompts give generic songs. The tips below cover what to specify.

Lead with the genre

Put the genre first. You can name an era or a region, or blend genres, as in “early 1990s boom-bap hip-hop” or “classical chamber music with dark electronic drone elements”.

Name the instruments, tempo, key, and mood

If you don’t name instruments, Lyria picks ones that suit the genre. To get a specific sound, say what plays and how it sounds:

Warm analog synth pads swelling beneath a dry, intimate acoustic guitar. Slow tempo around 72 BPM, in D minor. Melancholic and dreamy.

Shape the song

List the sections in order, as in [Intro] -> [Verse 1] -> [Chorus] -> [Verse 2] -> [Chorus] -> [Bridge] -> [Outro]. You can describe the energy in words too: “Build tension through the pre-chorus, then drop to silence before an explosive chorus.”

For tighter control, give each section a timestamp:

[0:00 - 0:10] Intro: a soft lo-fi beat and muffled vinyl crackle.
[0:10 - 0:30] Verse 1: add a warm Rhodes piano melody and gentle vocals about a rainy morning.
[0:30 - 0:50] Chorus: full band with upbeat drums and soaring synth leads.
[0:50 - 1:00] Outro: fade out with the piano melody alone.

Add your own lyrics

Keep the musical direction apart from the lyrics. Put the lyrics under a Lyrics: header and tag each section. Parentheses mark backing vocals and ad-libs.

Dreamy indie pop at 110 BPM with breathy female vocals.

Lyrics:

[Verse 1]
Early morning rain on the window pane
City lights wash away the pain

[Chorus]
We keep moving on (moving on)
Until the morning light

Or let Lyria write the lyrics

Give the story, the mood, or a key phrase. For dance tracks, ask for a short hook that repeats:

An upbeat dance-pop track with a repetitive, high-energy vocal hook: "Feel the rhythm all night long."

Describe the voice

Name the gender, the range, and the tone, as in “Female alto with a rich, warm, husky lower range.” You can also ask for a spoken intro, a whisper before the drop, or chopped vocal samples.

Make an instrumental

Lyria 3.5 sings by default. For an instrumental, add “Instrumental only, no vocals.”

Sing in another language

Lyria 3.5 sings in the language of your prompt. To get French lyrics, write the prompt in French.

Compose from images

Add images and say what the song should take from them, as in “An atmospheric ambient track inspired by the mood and colors in this image.” Google’s guide suggests thinking about who is in the image, where it is, and what is going on. Set the genre, instruments, and tempo in the prompt as usual.

Set the length

Ask for a length, as in “create a 2-minute song”, or end your timestamps where the song should stop. The length you ask for guides the song, but the result is not exact.

Try ideas on Lyria 3 first

Lyria 3 is faster and makes 30-second clips. Google suggests testing a prompt there before you generate the full song here.

Limitations

  • Each prediction makes a new song. You can’t edit or extend a song you already have, so change the prompt and run it again.
  • The same prompt and seed can give a different song each time.
  • There is no negative prompt. To leave something out, say so in the prompt, as in “no vocals”.

Safety and watermarking

  • Replicate checks the prompt and images for sexual content and graphic violence before sending them to Google.
  • Google blocks prompts that ask for a specific artist’s voice or for copyrighted lyrics. Describe the style, era, and instruments instead of naming an artist or a song.
  • Google’s filters can also block a harmless prompt, including some with long custom lyrics. If that happens, reword the prompt and try again.
  • Every song carries SynthID, Google’s watermark for AI-generated audio. You can’t hear it. To check a file, upload it to Gemini and ask if Google AI made it.

Learn more

Model created
Model updated