Lyria 3.5 in Gemini: AI music features and a practical production guide

What Google announced, and what it did not prove

Google announced Lyria 3.5 on September 4, 2026 and says the update improves vocals and arrangements. That claim should be attributed to Google because this article has not conducted a listening test. The announcement also describes ways to steer creation in Gemini, including genre descriptions, vocal versus instrumental direction, templates and short or long choices. Those controls are useful as a briefing surface, but they should not be confused with a guarantee that every prompt will produce a production-ready song.

DeepMind’s Lyria page says the system can generate tracks up to three minutes. That is an important capability boundary, but actual limits can vary by interface, account or product route. A user working in Gemini should check what that account currently offers, while a developer should consult the separate music-generation API documentation. Do not transfer an interface limit from one surface to another without verification. The same separation applies to terms, quotas and implementation details. Treat Gemini interaction and API integration as related but distinct workflows.

Start with the edit function, not an artist reference

A strong music brief begins with what the picture needs rather than a request to imitate a living or named artist. For a 30-second understated product film, define the job in seconds: leave space for dialogue at the opening, establish a restrained pulse, create one modest lift near the reveal and reach a clean ending that an editor can cut on. Describe the sound as sparse felt piano, soft percussion, controlled low-end and limited density. The creative target is 30 seconds because of the film, not because it represents Lyria’s maximum.

This functional language is more useful than vague adjectives such as ‘premium’ or ‘cinematic.’ Specify what must not happen: no crowded lead melody under dialogue, no abrupt genre switch, no oversized climax, no tail that forces the editor to fade awkwardly. If vocals are not needed, choose or request instrumental direction rather than leaving the issue ambiguous. If vocals are required, define their role and intelligibility without asking for a recognizable artist imitation. A good brief narrows decisions while leaving room for the system to compose.

Turn one brief into a controlled comparison

Generate several variants from the same core brief before rewriting the concept. Keep timing, instrumentation and narrative function stable so the differences are interpretable. Label each draft with a simple version code and note which setting or prompt phrase changed. The goal is not to produce dozens of tracks; it is to learn whether one controlled variation improves the fit. If every generation uses a different genre, tempo idea and structure, the review becomes taste browsing rather than an evaluation of the brief.

Compare the drafts at the same playback level. A louder file can feel more exciting even when its arrangement is less appropriate. Listen once for picture function, once for dialogue space, once for transitions and once for ending behavior. Do not claim that one variant has ‘better audio quality’ unless you have a defined technical measurement or listening protocol. For editorial review, use concrete observations such as ‘the percussion enters before the spoken line ends’ or ‘the final hit gives the editor a clear cut point.’ Those notes are actionable and reproducible.

Test the music where the audience will actually hear it

A production candidate should survive more than studio headphones. Check it on headphones for detail, a phone speaker for midrange balance and mono for compatibility. Use the same approximate playback level when comparing versions. The phone test is especially useful for sparse product films because low-frequency elements may disappear while a piano attack or vocal can become disproportionately prominent. Mono can reveal whether a key element loses impact when channels collapse. These checks do not require claiming a formal mastering test; they are practical editorial listening conditions.

Review the track against the actual picture, not only as standalone music. A composition can sound attractive alone yet compete with product sound effects, voice-over or pacing. Mark exact timestamps where the arrangement crowds dialogue, where a lift arrives too early or where the ending fails to align with the final frame. If the client is approving music, show the candidate inside the intended edit whenever possible. Approval of an isolated track and approval of a track in context are not the same production decision.

Rights and provenance need their own checklist

Do not describe a Lyria-generated track as automatically copyright-free, exclusive or commercially licensed unless the applicable terms and project-specific rights review support that conclusion. The documented facts do not establish those legal outcomes, and this article does not provide a legal conclusion. Before commercial use, check the current terms for the exact product surface, the intended territory and the planned form of distribution. Also consider any third-party rights implicated by the prompt, supplied material or requested creative references.

Maintain a provenance note for each selected asset. Record the tool or API route, generation date, core prompt, version identifier, whether external audio or lyrics were supplied, who selected the version and whether a terms or rights review was completed. Provenance is useful even when the project is low risk because it prevents the final WAV or other export from becoming an orphaned file with no production history. If the music is revised later, add a new version rather than overwriting the record. Traceability is part of creative operations, not only compliance.

Client approval should be a decision, not a vague reaction

Give the client two or three meaningful candidates at most, each presented against the same picture and playback conditions. Ask decision questions: does the track leave enough room for the spoken line, is the reveal lift restrained enough, and does the ending feel final without sounding theatrical? Avoid asking only ‘which one do you like?’ because that produces preference without production criteria. If none works, return to the brief and change one constraint rather than accumulating random variants.

Document approval in a simple record: version code, edit version, approver, date, requested changes and status. A generated track should follow the same change-control discipline as commissioned music. If a client requests a different vocal direction or a longer structure, treat that as a new revision with a new review. Do not silently substitute a regenerated file under the old version name. Clear versioning protects the editor, client and delivery team from mixing an approved track with a later experiment.

Build a practical delivery pack

The final delivery should contain more than a single audio file dropped into a chat. Include the approved master or project-appropriate audio export, a clearly named version, the picture edit it was approved against, timing notes for important transitions and a short provenance record. If the project requires stems or alternative lengths and the chosen workflow can produce them, list only the assets that were actually created; do not promise deliverables the product did not generate. Keep rights or terms documentation alongside the asset where the organization normally stores project records.

For the 30-second product-film example, the editor might receive one approved track, a note that dialogue occupies the opening seconds, a timestamp for the restrained lift, the intended end cut and one backup variant if the final picture changes slightly. That package turns generative music into a managed production asset. Lyria 3.5 may make generation easier according to Google’s launch claims, but the professional value comes from briefing, controlled comparison, context listening, provenance and approval. Generation is one stage; delivery quality depends on the workflow around it.

Use a prompt worksheet instead of rewriting from memory

For each generation round, keep a worksheet with fixed fields: picture duration, dialogue windows, musical role, desired instrumentation, density, vocal status, lift point, ending behavior and exclusions. Change only one or two fields between rounds. This makes feedback legible and reduces the chance that a promising result disappears because the next prompt changed everything. The worksheet is also useful for client communication because it translates subjective reactions into production constraints that can be revised deliberately.

Treat regenerated audio as a new asset

Every regeneration that changes structure, vocal direction or timing should receive a new version identifier and a new approval status. Keep rejected variants long enough to understand what was changed, but do not let them circulate as current masters. If the final edit changes, verify that the approved track still fits before extending or regenerating it. This simple change log protects provenance and prevents a late creative adjustment from bypassing the same listening and rights checks used for the original selection.