Lyria 2
Last reviewed
Jun 3, 2026
Sources
10 citations
Review status
Source-backed
Revision
v1 · 1,403 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
10 citations
Review status
Source-backed
Revision
v1 · 1,403 words
Add missing citations, update stale details, or suggest a clearer explanation.
Lyria 2 is a music generation model built by Google DeepMind that produces high-fidelity instrumental audio from text descriptions. It was announced on April 24, 2025 as part of an updated release of the company's Lyria music technology, which also introduced an interactive streaming variant called Lyria RealTime [1]. Google positioned Lyria 2 as a professional-grade text-to-music system, capable of rendering a range of genres and intricate arrangements while capturing the subtle differences between instruments and playing styles [1]. The model became generally available to enterprises through Google Cloud's Vertex AI platform on May 21, 2025, alongside the video model Veo 3 and the image model Imagen 4 [2]. As with the rest of the Lyria line, every clip it generates carries a SynthID watermark to mark the audio as AI-generated [1][2].
Lyria 2 is the second generation of a model family that Google DeepMind first unveiled on November 16, 2023, in partnership with YouTube [3]. That original Lyria was described at the time as the company's most advanced music generation model, and it launched with two applications. The first was Dream Track, an experiment on YouTube Shorts in which a small group of creators could type a topic, pick a participating artist from a carousel, and receive a 30-second soundtrack with an AI-generated voice, lyrics, and backing track in that artist's style [3]. The launch artists included Charlie Puth, T-Pain, Demi Lovato, John Legend, Sia, Charli XCX, Troye Sivan, Alec Benjamin, and Papoose [3]. The second was a set of Music AI tools developed through YouTube's Music AI Incubator, which later became the Music AI Sandbox [3][4].
The Lyria work drew on earlier Google research in music synthesis, most directly MusicLM, the text-to-music model introduced in early 2023. From the first release, Google committed to watermarking every Lyria output with SynthID, treating provenance marking as a default part of deployment rather than an optional add-on [3]. Lyria 2 inherited that approach and extended the model's fidelity and controllability.
Lyria 2 generates high-fidelity music and professional-grade audio outputs from text prompts, with finer control over instruments, tempo (BPM), and overall musical character than the original model offered [1][5]. Google describes the output as capturing nuance across genres and complex compositions [1].
On Vertex AI the model is exposed as lyria-002. Its output is instrumental-only audio delivered as 48 kHz stereo WAV clips of roughly 30 seconds [5][6]. The interface supports negative prompting, which lets a user describe elements to steer the model away from, and a single request can return more than one variation [5][6]. Generation passes through Google's safety stack, which the documentation describes as content safety filters, recitation checking, and artist intent checks meant to block harmful or inappropriate inputs and outputs [6].
| Attribute | Lyria 2 (lyria-002) |
|---|---|
| Developer | Google DeepMind |
| Announced | April 24, 2025 [1] |
| Vertex AI general availability | May 21, 2025 [2] |
| Output | Instrumental, 48 kHz stereo WAV, about 30 seconds [5][6] |
| Prompt controls | Text prompt, negative prompt, multiple samples [5][6] |
| Watermark | SynthID [2][6] |
The April 2025 announcement centered on a refreshed Music AI Sandbox, an experimental suite of tools aimed at working musicians, producers, and songwriters [1]. The Sandbox is built for tasks such as generating fresh instrumental ideas, crafting vocal arrangements, and breaking through creative blocks, and the April update folded Lyria 2 in as its underlying generation model [1]. Alongside the new features, Google widened access, opening the Sandbox to more creators in the United States through a waitlist sign-up [1]. The announcement featured early users including the artists The Range, Isabella Kensington, Adrie, and Sidecar Tommy [1].
The Sandbox traces back to the November 2023 Music AI tools, which had been limited to participants in YouTube's Music AI Incubator [3][4]. The 2025 relaunch broadened that footprint and paired the toolset with both Lyria 2 and the new Lyria RealTime model for interactive work [1].
Lyria RealTime is a distinct model variant announced in the same April 2025 release, designed for interactive, real-time music generation rather than fixed clips [1]. Where Lyria 2 renders a finished segment, Lyria RealTime produces a continuous stream of music that a user can steer, shape, and warp while it plays, mixing genres and blending styles moment by moment [1][7]. Google has compared the experience to jamming with a musician in a band [8].
Technically, the model generates a continuous stream of 48 kHz stereo audio in sequential chunks, an approach Google calls block autoregression that adapts the MusicLM architecture for live use [7][8]. Each chunk is conditioned on the previous audio and on a style embedding derived from the user's prompts, and the system keeps latency to a maximum of about two seconds between a control change and its audible effect [7][8]. Beyond blending text descriptors, the API exposes direct controls over tempo, key, note density, and spectral brightness, along with the ability to mute instrument groups such as drums or bass [7][8].
Google brought Lyria RealTime to developers as an experimental API at Google I/O 2025, on May 20, 2025, making it available through the Gemini API and Google AI Studio [9][10]. It is the model that powers MusicFX DJ, an experimental music app in Google Labs [9][10]. On the Gemini API the model identifier is models/lyria-realtime-exp; it generates instrumental music only and outputs 16-bit PCM audio at 48 kHz in stereo, with tempo selectable across roughly 60 to 200 BPM [10]. As with Lyria 2, output audio is always watermarked for identification under Google's responsible AI principles [10].
Lyria 2 reached general availability on Vertex AI on May 21, 2025, where it can be tested in Google Cloud's Media Studio and called programmatically through the model API [2][5]. The Vertex release grouped Lyria 2 with Veo 3 and Imagen 4 as a set of generative media tools for enterprise customers, all built with Google DeepMind and shipped with safety as a design principle [2]. Lyria RealTime, by contrast, is offered to developers as an experimental capability through the Gemini API and AI Studio rather than as a Vertex enterprise product [9][10].
Both models embed SynthID, Google DeepMind's imperceptible audio watermarking technology, into their outputs by default [2][6]. SynthID adds a signal that is inaudible to listeners and does not degrade the audio, yet remains detectable so that AI-generated music can be identified later [3]. Google has described the SynthID audio watermark as robust to common transformations such as MP3 and AAC compression. On Vertex AI the watermark is applied automatically to generated clips [6].
Lyria 2 marked Google's move from music generation as a closed YouTube experiment toward a commercially available service for businesses and developers. By placing the model on Vertex AI next to Veo 3 and Imagen 4, Google assembled a single generative media portfolio spanning music, video, and images, each watermarked with SynthID [2]. The accompanying Lyria RealTime model pushed in a different direction, treating music less as a file to download and more as a live, steerable performance, which opened uses such as interactive soundscapes for games and installations [1][8]. The pairing of a high-fidelity clip generator with a real-time streaming engine, both released together and both provenance-marked, set the template that later Lyria releases would build on. Google has continued to emphasize artist collaboration and responsible deployment, including artist intent checks and default watermarking, as it has broadened access to the technology [1][6].