Google Opens Lyria 3.5 to Developers, Generating Full Songs With Vocals

Google has pulled its top-tier music generation model out of its walled garden. Lyria 3.5 is now available in AI Studio, through the Gemini API, and inside the Gemini app, giving developers a programmatic path to full songs with vocals, lyrics, and instrumental arrangements. Previously, the model was only reachable through Google Flow Music, the company's consumer-facing creation platform.

Google promises more expressive singing, richer arrangements, and higher fidelity than Lyria 3. The API exposes two variants aimed at different production stages, both generating 44.1 kHz stereo audio from text prompts or images.

Two models, one API

• Lyria 3.5 Clip (lyria-3.5-clip-preview): fixed 30-second MP3 outputs for loops, previews, and prompt iteration.
• Lyria 3.5 Pro (lyria-3.5-pro-preview): multi-minute songs with distinct verses, choruses, and bridges, with duration controllable through the prompt.

Both run on the new Interactions API and accept multimodal inputs. Feed up to 10 images alongside a text prompt and the model composes music shaped by the visual mood. Calling it looks like any other Gemini SDK request:

from google import genai
client = genai.Client()
interaction = client.interactions.create(
    model="lyria-3.5-pro-preview",
    input="An epic cinematic orchestral piece. Solo piano intro, sweeping strings, massive climax.",
)
with open("song.mp3", "wb") as f:
    f.write(base64.b64decode(interaction.output_audio.data))

Four things that improved over Lyria 3

The 3.5 release is a quality pass rather than an architecture overhaul, targeting musicality, lyrics, vocals, and creative control:

• Richer melodic structures that sound more natural and complex.
• Generated lyrics follow prompts more accurately with clearer song structure.
• More realistic singing with improved pronunciation and emotional nuance, historically the dead giveaway on AI-generated tracks.
• Direct control over tempo, duration, key, and section timing via bracket tags like [Verse], [Chorus], and timestamped blocks like [0:00 - 0:10].

Scripting structure and lyrics

Developers get substantial prompt-level control. Supply your own lyrics with section tags, or script the song timeline down to ten-second windows specifying when instruments enter and when the mood shifts. Prompt in French and the vocals come out in French, with pronunciation adapted to the input language.

Responses are structured: each interaction returns a sequence of steps containing base64-encoded audio blocks alongside text blocks with the generated lyrics and a JSON breakdown of the song structure. That means you can render karaoke-style timelines without a second transcription pass.

Limits worth knowing before you build

• Single-turn only. No iterative editing. Tweaking a chorus means regenerating the whole track.
• Non-deterministic. The same prompt produces a different song each time. Seed control is not exposed.
• Safety filters block artist mimicry. Prompts requesting specific artist voices or copyrighted lyrics are rejected.
• SynthID watermarking is mandatory. Every track carries SynthID, Google's audio watermarking system, which embeds a machine-detectable identity so AI-generated music can be flagged as such.

The licensed-data angle

Google states Lyria was trained on licensed data, specifically materials YouTube and Google had rights to use. That matters because Suno and Udio are both facing copyright litigation from Sony Music. For anyone shipping a commercial product, a licensed-data claim combined with a mandatory watermarking pipeline carries a meaningfully different legal risk profile than the current market leaders.

Google also offers Lyria RealTime for streaming generation, completing a full music stack inside the same Gemini SDK developers already use for text, image, and video work. The legal and integration friction that made adding a soundtrack generator feel risky just got considerably smaller.