Lyria 3.5: Google strengthens AI music in Flow

Lyria, from technology demonstration to creative tool

Google DeepMind has announced the rollout of Lyria 3.5 in Google Flow Music, an environment presented as being intended for music creation. In its original post, entitled “We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control”, the lab highlights four areas: musicality, lyrics, vocals and finer creative controls.

The wording matters. The challenge is no longer simply to ask an artificial intelligence to produce a track from a descriptive sentence. Google DeepMind is seeking to establish Lyria 3.5 in a creative workflow where the user can better steer the outcome: decide on an intention, refine writing, intervene on the voice or guide the composition. The release does not publicly detail the exhaustive list of commands, the audio formats produced, the available durations, pricing terms or covered territories. But the emphasis on creative control reveals the direction being taken.

This announcement is part of an already lengthy history, on the very fast timescale of generative music. Google DeepMind unveiled the Lyria name in 2023, alongside several projects devoted to music generation. The group then presented the model as a high-quality music-generation system and linked it to experiments conducted with YouTube, notably around Dream Track for Shorts. This first stage already placed generative music in Google’s broader ecosystem: AI research, creation tools and the video distribution platform.

In 2024, Google DeepMind also introduced Music AI Sandbox, a series of experimental tools designed with the participation of artists, producers and songwriters. The lab cited Wyclef Jean, Marc Rebillet and Donald Harrison Jr. among the people who contributed to this effort. The idea was already to move beyond the automatic song generator and offer tools for ideation, transformation and production support.

Lyria 3.5 and Flow Music therefore extend a consistent trajectory: moving from a generation experience to a more integrated working layer, in which AI is intended to become an instrument for composition, arrangement or pre-production. This distinction is central for creators. A spectacular result obtained in a few seconds can impress in a demonstration; a genuinely usable tool must enable feedback, corrections, variations and alignment with an artistic intention.

The name Flow also evokes Google’s intention to bring together several generative capabilities in creative interfaces. Google launched Flow as an audiovisual filmmaking tool drawing in particular on its Veo, Imagen and Gemini models. With Flow Music, the group extends this vocabulary of creative “flow” to sound. The industrial logic is clear: in digital creation, image, video, dialogue, sound effects and music are no longer entirely separate stages. Creators are looking for tools capable of making them interact without multiplying exports, software applications and manual iterations.

DeepMind’s release does not, however, describe Flow Music as a conventional digital audio workstation. It does not make it possible to conclude that it replaces a sequencer, mixing software or a production environment such as those used in studios. It presents, first and foremost, a model and a creative space. This caution is necessary in a market where promises around AI music are often broader than the technical information actually published.

What Google is specifically announcing with Lyria 3.5

The main point is simple: Lyria 3.5 is launching in Google Flow Music. Google DeepMind highlights advances in four areas corresponding to the most visible weaknesses of earlier generations of AI music: musical coherence, lyric quality, the realism or control of vocals, and the ability to steer the result.

Musicality, a broader criterion than sound quality

In the vocabulary of generative AI, “musicality” does not refer only to better audio definition. It encompasses a track’s ability to hold together over time: maintain a credible rhythm, build a harmonic progression, preserve consistent instrumentation, provide transitions and give the impression that the sections belong to the same intention. For a user, the difference is concrete. A generation may have a pleasant timbre for a few seconds while collapsing as soon as one examines the structure, chord changes or evolution of energy.

Google DeepMind therefore says that Lyria 3.5 is making progress in this area, without providing in its announcement the test measurements, standardized comparative excerpts or external evaluations that would make it possible to quantify the gain. The post also does not publish a common benchmark with competing models. The promise should therefore be considered for what it is: a product statement issued by its publisher, whose practical scope will depend on the results obtained by users in Flow Music.

The question of musicality nevertheless remains strategic. Creators do not assess a generated track solely by how immediately convincing it is. They seek musical material that can fit into a video, serve as a demo, feed a writing session or accompany a campaign. In these uses, the level of control over structure is often worth more than a surprise effect. Music must support a narrative, not distract from voice-over, prepare a change of shot or respect a brand identity. The improvement announced by DeepMind addresses this requirement.

Lyrics and vocals: the most sensitive areas

Lyrics are explicitly among the announced advances. This is an area where music models have often encountered two difficulties: the intelligibility of sung text and its alignment with the user’s request. A phrase may be pronounced indistinctly, truncated, repeated without justification or transformed over the course of generation. Setting text to music also involves prosody, pacing, rhymes, breathing and articulation. Promising advances in lyrics therefore amounts to claiming better management of a linguistic and musical dimension simultaneously.

For the French-speaking market, this issue is particularly closely watched. French song often relies on understanding the text, the placement of consonants, liaisons, meter and the relationship between the voice and words. Google DeepMind’s release does not specify which languages are supported, nor whether French receives special treatment. It would therefore not be rigorous to infer a given level of performance in French. But the attention paid to lyrics makes this test unavoidable as soon as French creators can access the product under their respective terms of use.

The same reasoning applies to vocals. Google announces progress in this area, but does not specify, in the information provided, the types of voices, configuration options, safeguards against imitating identifiable people or content provenance mechanisms. Yet voice is both an aesthetic component and a particularly sensitive legal area. It contributes to a performer’s identity; it can also suggest the participation of a person who did not contribute.

Vocal quality is also a criterion of technical maturity. The human voice quickly exposes inconsistencies: imprecise attacks, unstable vowels, artificial phrasing, ambiguous diction, poorly placed breaths, unexpected changes in register or sonic personality. A music-creation tool claiming an improvement in vocals is therefore not measured solely by its ability to produce an appealing timbre. It is measured by its ability to maintain a credible performance throughout a composition, and to let the user steer it without creating breaks.

Creative control as the core promise

The fourth area is probably the most structuring: Google DeepMind emphasizes finer creative controls. The release does not detail the interface or the exact taxonomy of these controls. They should therefore not be assigned specific functions that have not been announced. However, the strategic signal is clear: Google wants Flow Music to be seen not as a black box delivering a finished song, but as an environment in which the creator guides the AI.

This ambition responds to a frequent criticism of generative music. When a user formulates a vague instruction, they receive a result that may be technically impressive but difficult to reproduce, modify or connect to an existing project. The more the system lets the user act on the variables that matter, the more it can fit into a professional process. In music, these variables may concern mood, style, energy, writing, instrumentation, duration or the track’s evolution. DeepMind’s announcement does not list what is actually accessible in Lyria 3.5; it nevertheless places user control at the center of its message.

This difference is also cultural. For a musician, “creating” does not necessarily mean producing every note by hand, but being able to own choices: deciding that a chorus should arrive later, that the voice should fade away, that the rhythm should become more tense or that the arrangement should make room for dialogue. Artistic direction is precisely the art of holding these decisions together. Flow Music will be judged on its ability to make these choices revisable and intelligible, rather than on the sheer volume of music it can generate.

A market where raw generation is no longer enough

Lyria 3.5 arrives in an already crowded competition. Suno and Udio have helped popularize among the general public the idea that text could give rise to a complete song, with instrumentation, structure and singing. Stable Audio, developed by Stability AI, has in turn established itself in the audio-generation landscape, with an approach covering in particular the creation of sounds and music. These products are not identical in their interfaces, commercial terms, formats or orientations, but they all compete to redefine the entry point for sound creation.

The useful comparison is not limited to the perceived quality of an excerpt. Each platform is trying to solve an equation combining four dimensions: accessibility, control over the result, legal safety and integration into real production workflows. A very simple interface can attract a wide audience, but frustrate professionals if it does not allow iteration. Conversely, a highly configurable system may require expertise that puts off occasional users. An extensive catalog of styles can appeal, but it also raises more acute questions when it appears to approach existing artists.

On this board, Google has a potential advantage: integration. DeepMind belongs to a group that operates distribution platforms, productivity tools, creative environments and multimodal models. Google can therefore think about music in a broader chain: music for a video, a sequence for a campaign, a sonic identity for social content or accompaniment for a presentation. The existence of Flow on the audiovisual side makes this coherence especially visible.

But integration does not guarantee adoption. Professional creators work with established software, sound libraries, sound engineers, composers, publishers and approval processes. They need to know what they can use, modify, distribute, monetize and archive. At this stage, the information summarized by DeepMind’s release does not specify the status of productions generated with Lyria 3.5, the rights granted to users or the restrictions applicable to certain commercial uses.

This lack of detail does not mean that no framework exists. It means that it cannot be established from the announcement itself. For agencies, labels, independent producers and media organizations, this point is at least as important as improvements in musicality. In a production commissioned by a brand or intended for paid distribution, traceability and contractual clarity directly determine a tool’s value.

Competition is also playing out over iteration speed. Users of generative music do not always seek the final track from the first request. They test a direction, compare several variants, retrieve a chord idea, a drum color, a texture or a starting point for a session. Systems that facilitate feedback and modifications have a chance to become everyday tools. Those that produce only isolated results remain demo generators, sometimes useful, but less deeply integrated into creative work.

DeepMind’s positioning around “creative controls” is therefore a direct response to this evolution. After the era of the spectacular prompt, the battle is shifting toward editing and intention. Users do not merely want to describe music; they want to be able to say what must change when the first version does not work. This is where Flow Music’s promise will have to be tested against actual use.

Data, rights and trust: the decisive ground for the industry

Generative music cannot be analyzed without addressing its development conditions. Models learn from very large datasets, and their capabilities have sparked global debates over copyright, compensation, transparency and the place of artists. In 2024, record labels represented by the Recording Industry Association of America brought proceedings against Suno and Udio, alleging in particular copyright infringement. The companies concerned challenged the allegations and defended their respective positions. These disputes illustrate a broader conflict that extends far beyond the two companies.

Major labels, collective management organizations, publishers and platforms are seeking to define the conditions under which generative technologies can use works, recordings or protected elements. At the same time, several industry players have explored collaborations with AI companies. The market therefore cannot be reduced to a binary opposition between technology and rights holders: it concerns consent, licensing, compensation, data governance and artists’ ability to choose how their work is used.

The release on Lyria 3.5 does not provide, in the information given, details on the model’s training data. Nor does it detail any licensing agreements, opt-out options, revenue-sharing mechanisms, voice policies or detection tools. These subjects should not be filled in with assumptions. On the contrary, they are the questions Google will have to answer clearly if it wants to convince the most exposed professionals.

Google DeepMind does not approach generative music without experience in dialogue with the creative industry. The presentation of Music AI Sandbox already emphasized an approach built with artists and professionals. The Dream Track experiment on YouTube also involved artists who agreed to take part in the experiment. This establishes a precedent in method: working with creators rather than placing them only in front of a finished product. But this precedent does not by itself allow the terms of Lyria 3.5 in Flow Music to be inferred.

For a musician, trust is not limited to the abstract question of data. It concerns the effect on their activity. Can a voice be imitated? Can a personal style be used in a way close enough to mislead the public? Can works released with AI be identified? Will an author be paid if their work contributes to the value of a system? The answers fall under law, contracts, design choices and platform policies alike.

In Europe, the regulatory framework adds another layer. The European Union’s AI Act entered into force in 2024 and establishes progressive obligations for AI actors, including around general-purpose AI models. The regulation does not by itself settle all debates related to music copyright, but it strengthens the demand for transparency already running through the sector. For providers active in the European market, compliance and documentation are becoming differentiating factors just as important as a model’s quality.

France is particularly attentive to these issues, given its production ecosystem, its collective management organizations, the place of the French language and the importance attached to copyright. Film composers, post-production studios, independent labels, web creators and agencies do not all have the same needs. They nevertheless share one requirement: knowing the exact scope of what they can do with generated music, and understanding the residual risks.

From this perspective, Lyria 3.5 will not be assessed only as a model. It will be assessed as a proposition of trust. Google DeepMind can score points if Flow Music comes with clear information on rights, restrictions, provenance and avenues for redress. Conversely, even a major improvement in musicality or vocals will not be enough to dispel the reservations of organizations that must manage catalogs, contracts or national campaigns.

What Flow Music could change for French creators

For individual creators, a tool like Flow Music could first serve to speed up the early stages of a project. A videographer may seek a sonic direction for an edit. A podcaster may test a musical identity. A singer-songwriter may use a generation as a discussion point in a working session. A small agency may explore several atmospheres before commissioning an original composition. These uses are not equivalent to the automatic publication of a generated track, and this is precisely where the controls announced by DeepMind become relevant.

The promise of AI-assisted artistic direction could reduce the time spent on demos and sound research. In short formats, social content or internal communications, speed is often decisive. Creating an atmosphere consistent with an image, voice-over or brand message is work that can be costly when it requires extensive research and licensing. If Flow Music allows users to guide results precisely, it could become a particularly attractive ideation tool.

However, AI does not automatically solve questions of taste or identity. A usable track is rarely merely an assembly of genre references. It is part of a project, an audience, a budget and sometimes an artist’s story. Artistic direction also requires knowing how to reject an overly obvious solution, preserve a silence, take a risk or retain an expressive imperfection. Nothing in DeepMind’s announcement claims to eliminate this human role. The ambition instead appears to be to make the system more manageable in this dialogue.

French-speaking creators will in particular have to observe the actual treatment of sung and spoken French. A tool’s value is not the same depending on whether one is making instrumental music, a song with lyrics or an advertisement in which words must be perfectly understood. The lyric quality announced by Google is therefore a point to verify in different situations: accents, proper names, idiomatic expressions, rap flow, lyric-driven song, choral voices and the possible coexistence of several languages. The release gives no answer to these questions, but it makes lyrics one of Lyria 3.5’s areas of work.

For professionals, the question of integration with existing tools will be just as decisive. Studios work through versions, approvals, deliveries, stems, mixes and synchronization rights. Yet the announcement does not specify which audio elements Flow Music allows users to export, isolate or rework. Without these details, it would be premature to present the product as a direct alternative to traditional production chains. It can, however, be added to those chains as a research, previsualization or prototyping tool.

This distinction between replacement and complement is essential to the French debate. AI music is sometimes described as a uniform threat to every profession. In practice, its effects will differ. Low-budget or high-volume productions may be the first to shift toward automated tools. Other areas, such as custom music for film, video games, live performance or the recording artist, rely on creative and contractual relationships that generation does not mechanically reproduce. Google’s model could accelerate certain uses while creating new demand for supervision, selection and finalization.

There is also a question of concentration. When creation, distribution, advertising and search tools belong to very large technology groups, creators potentially gain in fluidity but may lose autonomy if formats, monetization rules and access conditions change. Google has considerable reach in digital uses. This position can facilitate adoption of Flow Music, but it also requires particular vigilance regarding interoperability, portability of creations and clarity of terms of use.

The next step: proving that control is real

Lyria 3.5 opens a phase in which demonstrating quality will have to give way to demonstrating control. Google DeepMind says it has made progress in musicality, lyrics and vocals; these are important criteria, but they take on meaning only if the user can employ them predictably. A creator must be able to understand why a result was obtained, how to make it evolve and to what extent a new request will retain the elements they like.

This requirement could become the main differentiating factor against Suno, Udio and Stable Audio. Competition will not be played out solely on the ability to produce a complete track from a sentence. It will concern the ability to turn a draft into an artistic decision: retain a motif, change an intention, adjust text, obtain a variation compatible with the project and document what can subsequently be used. The platforms that make this journey understandable will have an advantage among professionals.

For Google, the challenge is also to turn its multimodal expertise into a coherent creative experience. Music is rarely consumed or produced alone. It accompanies images, games, advertisements, videos, podcasts and performances. The existence of Flow on the video side gives the company an opportunity to connect these uses. But this integration will have to avoid reducing music to a simple automatic accompaniment. Creators will adopt a tool over the long term if it leaves them an identifiable place in decisions, not if it turns the soundtrack into an interchangeable variable.

The next stage will therefore depend on the details Google DeepMind chooses to make public: Flow Music’s availability, the scope of controls, languages, rules relating to voices, user rights, commercial terms and transparency guarantees. This information will determine Lyria 3.5’s concrete reach in studios and among independent creators.

Over the longer term, the announcement confirms a market transformation: generative music is moving toward artistic steering interfaces rather than instant production alone. If Google succeeds in combining quality, control and a framework of trust, Flow Music could become an important building block of AI-assisted audiovisual creation. If these three dimensions remain disconnected, the tool risks joining the long list of impressive generators that are difficult to integrate sustainably into professional music practice, particularly in a French-speaking and European space where law, language and recognition of creators remain inseparable from innovation.