Lyria 3.5 Stem Export 2026: Copy-Paste Prompt Recipes

Google shipped Lyria 3.5 inside Gemini in early September 2026, and the headline feature is the one musicians have been begging for since the first AI music demos: separated instrument stems you can drag straight into a DAW. Longer generations matter too — you can finally get a full two-and-a-half-minute bed instead of a 30-second loop that dies mid-phrase. But there’s a catch nobody is writing about: Lyria 3.5 prompts are structured prompts. Type “chill lofi beat” and you’ll get four stems that all sound like the same mud smeared across four files. Type the right BPM, key, section markers, and negative-instrument tags and you get a mixable multitrack. This guide is the prompt syntax, copy-paste ready.

What’s actually new in Lyria 3.5 prompts

Lyria 3.5 is Google’s third-generation music model, and it landed in the Gemini app and the Gemini API in the first week of September 2026. Three things changed materially from Lyria 2. First, generation length: the practical ceiling moved from roughly 30 seconds to about 150 seconds of coherent output, with structure that holds across the whole span rather than looping a bar and hoping. Second, and the reason you’re reading this, stem export. A single generation now returns separated tracks — typically drums, bass, harmonic/chords, and lead/melody — as individual audio files rather than one flattened mix.

Third, and least documented: the prompt parser got far more literal. Earlier Lyria builds treated your text as a vibe description and did what they wanted with tempo. Lyria 3.5 respects explicit numeric BPM, respects key signatures written in standard notation, and respects exclusion syntax. Write no vocals, no cymbals and it holds those out of the stems. That’s the difference between a stem pack you can mix and four files that bleed into each other.

The stem separation is not a post-hoc source-separation pass bolted onto a stereo render. Lyria 3.5 generates the parts as parts, which is why the isolation is clean instead of the smeared, phasey artifacts you get from running Demucs over a finished mix. That distinction matters if you plan to process these — you can compress the drum bus hard, sidechain the bass, and re-EQ the pad without the artifacts fighting you. Gemini music generation went from “interesting demo” to “usable asset pipeline” in one release, and almost nobody has published the syntax.

Why it matters

• Royalty-free AI background music becomes a solved problem for video creators. A 150-second bed with stems means you can duck the melody under voiceover and leave the drums and bass carrying the energy — the standard mix move that separates amateur YouTube audio from professional.
• Stems kill the “AI music is unusable” objection. The complaint was never that AI music sounded bad; it was that you couldn’t fix it. A flattened mix is take-it-or-leave-it. Four stems is a starting point you can produce.
• Prompt structure is now a real skill with real leverage. The gap between a lazy prompt and a structured one is larger in Lyria 3.5 than in any image model. AI music prompt structure is closer to writing a chart than writing a caption.
• Podcast and course producers get consistent sonic branding. Lock a key, BPM, and instrument list, vary only the section markers, and you can generate intro, outro, transition, and bed variants that genuinely sound like the same piece of music.
• Game and app developers get adaptive layers for free. Stems are exactly what an adaptive audio system wants: mute the lead in the menu, bring in drums during combat. That used to require hiring a composer who understood interactive scoring.
• Licensing clarity is improving but is not settled. Google positions Lyria output as usable commercially with SynthID watermarking embedded, which resolves the provenance question but not every platform’s policy about AI-generated audio. Check the terms for your distribution channel, not just Google’s.

How to use Lyria 3.5 stem export today

1. Start from the four-block prompt skeleton. Every reliable Lyria 3.5 prompt has the same four blocks in the same order: technical spec, instrumentation, structure, exclusions. The model weights early tokens more heavily, so technical spec goes first — the single biggest change from how you prompted Lyria 2.

   [SPEC] 92 BPM, key of D minor, 4/4, straight (non-swung) feel, 140 seconds
   [INSTRUMENTS] warm upright bass, brushed drum kit, Rhodes electric piano, muted trumpet lead
   [STRUCTURE] 0:00-0:16 intro (Rhodes + bass only) | 0:16-0:56 A section (full band, trumpet states melody) | 0:56-1:36 B section (trumpet drops out, Rhodes solos) | 1:36-2:20 A section return, ritardando final 4 bars
   [EXCLUDE] no vocals, no synth pads, no cymbal crashes, no sidechain pumping, no risers

2. Write the BPM as a number, not a word. “Mid-tempo” gets you anything from 70 to 110. 92 BPM gets you 92. This matters enormously for stem export: if you plan to layer a second generation on top of the first, the grids have to line up. Lock the number and reuse it.

3. Use timestamp section markers, not adjectives. “Builds to a climax” is a wish. A timestamped arrangement map is an instruction. The pipe-delimited format above works most consistently — the model parses the ranges and places the transitions near them (expect ±2 seconds, not sample-accurate).

4. Write negative-instrument tags aggressively. This is the highest-leverage line in the whole prompt and the one everyone skips. Without exclusions, Lyria 3.5 defaults toward dense, full-spectrum arrangements, which produces stems that overlap in frequency and refuse to sit in a mix. Every instrument you exclude gives the remaining stems room.

   [EXCLUDE] no vocals, no vocal chops, no choir, no strings,
   no white-noise risers, no reverse cymbals, no 808 slides,
   no tempo changes, no key changes, no fade-out ending

5. Request stems explicitly and name them. In the Gemini app, the stem toggle sits in the generation options panel. Via the API you request them in the call, and naming the split you want improves separation quality noticeably.

   Generate with separated stems. Split into exactly four tracks:
   (1) drums and percussion only
   (2) bass only
   (3) harmonic/chordal instruments only
   (4) lead melody only
   Do not duplicate any instrument across stems.

6. Call it from the API when you need batches. The app is fine for one-offs; the API is how you generate twelve variants of a course intro while you do something else. A minimal Python call:

   from google import genai
   
   client = genai.Client(api_key="YOUR_KEY")
   
   PROMPT = """
   [SPEC] 120 BPM, key of F major, 4/4, 90 seconds
   [INSTRUMENTS] plucked nylon guitar, soft kick and shaker, warm sub bass, glass marimba lead
   [STRUCTURE] 0:00-0:12 intro (guitar alone) | 0:12-0:50 main theme (full arrangement)
   | 0:50-1:30 breakdown then resolve, clean stop on downbeat
   [EXCLUDE] no vocals, no brass, no distortion, no fade-out
   """
   
   response = client.models.generate_music(
       model="lyria-3.5",
       prompt=PROMPT,
       config={
           "stems": True,
           "duration_seconds": 90,
           "seed": 44821,
       },
   )
   
   for stem in response.stems:
       with open(f"course_intro_{stem.name}.wav", "wb") as f:
           f.write(stem.audio_bytes)

7. Pin the seed once you find a keeper. Seeds are how you get variations that belong to the same family instead of four unrelated tracks. Lock the seed, change one block of the prompt, regenerate. Change the structure block to get an outro that matches your intro. Change the instruments block to get a sparse version of the same composition.

8. Import to your DAW at the declared BPM. Set your project tempo to the BPM you prompted before you drag the files in, then drop all four stems at bar 1. They arrive phase-aligned with each other because they came from one generation. Then do the production work Lyria won’t do for you: high-pass everything except the bass and kick around 80–100 Hz, and carve 2–4 kHz out of the harmonic stem so a voiceover has somewhere to live.

9. Run a loop-point check before you commit. If the bed has to loop under a long video, prompt for a clean stop rather than a fade-out (note the exclusion above), then trim to the nearest bar. A generation that fades cannot be looped; a generation that stops on a downbeat can.

How it compares

Tool	Stem export	Max length	Prompt control	Best for
Lyria 3.5 (Gemini)	Native, generated as separate parts	~150 seconds	Strong: numeric BPM, key, timestamped sections, exclusions	Instrumental beds, adaptive game audio, DAW workflows
Suno	Yes, post-generation separation	Multi-minute, full songs	Moderate: style tags plus lyrics, looser tempo adherence	Full vocal songs, complete tracks
Udio	Yes, on paid tiers	Extendable in segments	Moderate, with strong extend/inpaint editing	Iterative song building and remixing
Stable Audio	Limited	Several minutes	Good on tempo and duration; instrumental-focused	Sound design, loops, SFX
Traditional stock libraries	Sometimes, per-track	Any	None — you search, you don’t prompt	Guaranteed-clearance licensing

The honest positioning: if you want a song with vocals, Suno is still ahead. If you want an instrumental bed you’re going to mix yourself, Lyria 3.5 in Gemini is now the strongest option, and prompt adherence is the reason. Nothing else takes a timestamped arrangement map this literally.

What’s next

The obvious near-term gap is stem count. Four stems covers most video work and falls well short of real production — anyone who has mixed a live band wants the kick and snare on separate tracks, not a single “drums” bounce. Expect pressure on Google to expand the split, and expect the prompt syntax for requesting a custom stem layout to be one of the first things that changes. If you’re building tooling on this, keep the stem-naming block in your prompt template isolated so you can swap it.

The second thing to watch is length. 150 seconds covers a YouTube intro, a podcast bed, and a game loop, but not a three-and-a-half-minute track. The likely path is a stitching or continuation feature — generate a section, then extend it with the same seed and key — rather than a straight jump to five-minute single-shot generation. Udio already proved the extend-based approach works; whoever ships coherent, key-locked continuation with stems intact wins this category.

Finally, watch the rights layer. SynthID watermarking is embedded in Lyria output, which is genuinely useful for provenance and will probably become the mechanism platforms use to enforce AI-audio disclosure policies. That’s a feature, not a threat, if you’re doing legitimate work — but it does mean “nobody will know it’s AI” was never the plan and isn’t a viable strategy. Build on the assumption that your AI music stem export is detectable and labeled, and make the music good enough that it doesn’t matter.

Frequently Asked Questions

Do I need a paid Gemini plan to use Lyria 3.5 stem export?

Basic Lyria 3.5 generation is available on free Gemini tiers with usage limits; stem export and longer durations have generally been gated behind paid tiers, and the specifics shift. Check your current plan’s feature list before building a workflow around it — and if you’re generating in volume, the API costs less per track than upgrading purely for the app UI.

Can I use Lyria 3.5 music commercially in monetized videos?

Google’s terms permit commercial use of generated output, and SynthID watermarking is embedded for provenance. The caveat is downstream: YouTube, Spotify, and stock marketplaces each have their own AI-content disclosure policies that are evolving independently of Google’s license. Read your distribution platform’s rules, not just the model’s.

Why do my stems sound like the same instrument bleeding across all four tracks?

Almost always a missing exclusion block. Without negative-instrument tags, Lyria 3.5 fills the arrangement densely and instruments overlap in register, so the split has nothing clean to separate. Add [EXCLUDE] with five to eight specific instruments you don’t want, and name each stem’s contents explicitly in your stem request.

Does Lyria 3.5 actually respect the BPM I write?

Yes, when you write it as a number followed by “BPM” in the first line of the prompt. Adjectives like “upbeat” or “slow” get interpreted loosely. Put the numeric spec first — the parser weights early tokens more heavily, and this is the most reliable fix for tempo drift between generations.

Can I get vocals on a separate stem?

Lyria 3.5’s stem export is oriented toward instrumental splits, and vocal generation is deliberately constrained compared to Suno or Udio. For instrumental beds that’s fine and arguably an advantage — cleaner separation. If your project needs a lead vocal stem, generate the instrumental in Lyria and track or source the vocal separately.

How do I make multiple tracks that sound like the same project?

Lock three things across every generation: the seed, the key, and the BPM. Then vary only the structure block to produce an intro, a bed, a transition sting, and an outro. This is the most useful technique in this entire Lyria 3.5 tutorial, and it’s how you get a sonic identity instead of a folder of unrelated clips.