Hear Lyria 3.5
Lo-fi study beat
Instrumental, Rhodes and vinyl
Cinematic build
Orchestral, piano to full brass
Indie pop, sung
Vocals, written lyrics
Dark trap beat
Instrumental, 808s in D minor
French chanson
Vocals, prompted in French
Chiptune loop
Instrumental, retro 8-bit
What Lyria 3.5 is good at
Songs that hold their shape
The model plans the arrangement before it renders anything, so a two-minute track moves through an intro, verses, and a chorus instead of looping one idea. Write a timeline in and it builds to your marks.
Lyrics you wrote, sung back
Paste your verses under their section tags and keep them clear of your musical direction. The model performs the words rather than describing them, and hands back the lyrics as text when it writes them itself.
Scores built from your footage
Attach up to ten stills and the model composes to the mood, palette, and setting it reads in them. Feeding it frames from the timeline gets a cue that matches the grade rather than your description of it.
Instrumental beds with room to talk over
Ask for instrumental only and the arrangement leaves space where the dialogue sits. The usual route for game audio, explainer backing, and podcast themes.
A vocal in any language
The prompt language sets the lyric language, and the model adapts pronunciation to match. Write the whole brief in the target language, because an English prompt asking for a Spanish song is still an English prompt.
Cheap drafts before the real render
Lyria 3 Clip returns a fixed 30 seconds fast, so the prompt gets iterated there. Since generation is single-turn and non-deterministic, cheap drafting is what makes the full render land first time.
How to write a Lyria 3.5 prompt
Vague prompts produce generic music. Whatever you leave unsaid, the model renders as an average of everything it knows, so each thing you name is a decision you take back.
| Name this | Why it matters | Example |
|---|---|---|
| Genre | Sets the whole arrangement | Lo-fi hip hop, cinematic orchestral |
| Instruments | The exact timbre, not a family | Fender Rhodes, slide guitar, TR-808 |
| Tempo | A number beats an adjective | 85 BPM, slow around 70 BPM |
| Key | Fixes the mood harmonically | In G major, in D minor |
| Mood | Colours the performance | Nostalgic, aggressive, ethereal |
| Structure | Section tags or a timeline | [Verse] [Chorus], or timestamps |
Naming an exact instrument is worth more than naming a genre. "Fender Rhodes" gets a specific timbre; "keyboard" gets the average of every keyboard.
A 30-second lo-fi hip hop beat with dusty vinyl crackle, mellow Rhodes piano chords, a slow boom-bap drum pattern at 85 BPM, and a jazzy upright bass line. Instrumental only.
How to control Lyria 3.5 song structure
Once the sound is right, decide the shape. Two levels of control, and most people never reach the second.
Section tags mark the parts: [Intro], [Verse], [Chorus], [Bridge], [Outro]. Enough when you only care about the order.
Timestamps set the clock. Write a timeline and the model builds to your marks, which also fixes the duration.
[0:00 - 0:10] Intro: Begin with a soft lo-fi beat and muffled vinyl crackle. [0:10 - 0:30] Verse 1: Add a warm Fender Rhodes piano melody and gentle vocals singing about a rainy morning. [0:30 - 0:50] Chorus: Full band with upbeat drums and soaring synth leads. The lyrics are hopeful and uplifting. [0:50 - 1:00] Outro: Fade out with the piano melody alone.
This is what makes the model usable against picture: pull the in and out points from your edit, write them in as sections, and the swell lands on the cut. Without timestamps, ask in words instead, as in "create a 2-minute song".
How to use your own lyrics in Lyria 3.5
Paste the lyrics and tag the sections. One rule decides whether the model sings your words or paraphrases them: keep the lyrics separate from the musical direction. Direction first, then the lyrics as their own block.
Create a dreamy indie pop song with the following lyrics: [Verse 1] Walking through the neon glow, city lights reflect below, every shadow tells a story, every corner, fading glory. [Chorus] We are the echoes in the night, burning brighter than the light, hold on tight, don't let me go, we are the echoes down below.
Leave the lyrics out and the model writes them, then returns them as text beside the audio. Useful even when you plan to write your own: generate once, then edit what it chose instead of starting from a blank page. For anything dialogue sits under, say "instrumental only, no vocals".
Lyria 3.5 image input and multilingual lyrics
Images can set the score. Attach up to ten stills and the model composes to the mood, palette, and setting it reads in them. Feed it frames from your own timeline and the music matches the grade rather than your description of it.
The prompt language sets the lyric language. Write the brief in French and the vocal comes back French, pronunciation included. The catch: an English prompt asking for "a song in Spanish" is still an English prompt, and usually returns English. Write the whole thing in the target language.
Crée une chanson pop romantique en français sur un coucher de soleil à Paris. Utilise du piano et de la guitare acoustique.
Lyria 3.5 workflow and limits
Draft on Lyria 3 Clip, which returns a fixed 30 seconds fast, then render the finished prompt on Lyria 3.5 for the full song. Ask for WAV instead of MP3 when the track is going into an edit.
That order exists because of the four constraints below. Cheap drafts are the whole game.
| Limit | What it means | What to do |
|---|---|---|
| Single-turn | No "make the chorus bigger" follow-up | Change the prompt, generate again |
| Non-deterministic | The same prompt never repeats a track | Download anything you like immediately |
| Safety filters | Artist names and copyrighted lyrics are refused | Describe the voice, do not name the singer |
| SynthID watermark | Every track is marked as AI-generated | Nothing, it is inaudible |
FAQs
Two ways. Say it in words, as in "create a 2-minute song", or write a timestamped timeline and let the last mark set the end. Timestamps are the more precise of the two because they also decide what plays where. Lyria 3 Clip ignores both: it always returns exactly 30 seconds.
Clip is a fixed 30-second model built for fast iteration, and it returns MP3. Lyria 3.5 generates full-length songs of around two minutes with distinct verses, choruses, and bridges, and can also return WAV. The recommended workflow uses both: draft the prompt on Clip, then render the keeper on 3.5.
Yes. Include them in the prompt and mark the sections with [Verse], [Chorus], and [Bridge] tags. Keep the lyrics as their own block, separate from your musical direction, or the model tends to treat your words as a description of the song rather than the words to sing.
Ask for it in the prompt: "instrumental only, no vocals". It works on both models and is the standard approach for game audio, background beds, and any track that has to sit under dialogue.
Almost always one of two reasons. The prompt asked for a specific artist's voice or style by name, or it included copyrighted lyrics. Both are refused by the safety filters. Rewrite the prompt to describe the qualities you want, such as the timbre, the accent, and the delivery, rather than naming the artist who has them.
Not through a follow-up prompt. Music generation is single-turn, so there is no iterative refinement of an existing track. Change the prompt and generate again. Because results also vary between runs, download anything you want to keep as soon as you have it.
44.1 kHz stereo, the same sample rate as a CD master. MP3 is the default output, and Lyria 3.5 can also return WAV when you need the uncompressed file for an edit or a mix.
Yes. Attach up to ten images with your text prompt and the model composes to the mood, palette, and setting it reads in them. It is the most reliable way to make a score match footage, because the stills carry information your description would have to approximate.