Lyria 3.5

Google's music generation model, announced 2026-07-29 and shipping in Flow Music (flowmusic.google), a Google Labs product for making songs from prompts. Google's entry into the audio-music-generation branch alongside suno, udio and stable-audio — and the first song-generator in this wiki from a frontier text-LLM lab (google, llm-providers-wiki, cross-wiki).

What the announcement claims

Four improvements over the prior Lyria, all stated without numbers:

• Musicality — "richer, more complex melodic structures that sound more natural."
• Lyrics — "higher quality lyrics with improved prompt adherence and structural awareness."
• Vocals — "more realistic and emotionally nuanced vocals, plus improved pronunciation."
• Creative control — easier control of tempo and duration of the output.

A 2:37 Google Labs video demos one generated track.

What it does not say

The post is a product announcement, not a model card. It gives no architecture, parameter count, training data, output length or sample rate, generation latency, pricing or access tier, and no benchmark or listening-test result — so Lyria 3.5 cannot be placed against suno's Elo (~1293) or any other point in tts-benchmarks-style rankings. It also names no baseline version: "advancements across musicality, lyrics, and vocal quality" is measured against an unnamed predecessor.

Two absences matter more than the missing specs. Both are live axes in this wiki:

• No watermarking claim. Google's own gemini-live-3-5-translate announcement led with SynthID; this one does not mention it, or any provenance marking on generated audio. Whether Lyria 3.5 output carries SynthID is unstated, not denied — see audio-deepfake. Corrected 2026-10-01: the July post is still silent, but the Gemini API music-generation guide now states that "All generated audio includes a SynthID audio watermark". The claim has a vendor source. Nobody outside Google has tested whether the mark survives re-encoding.
• No rights or training-data statement. ai-music-copyright is the dominant axis in this branch (suno in litigation, udio licensed post-UMG settlement), and the announcement takes no position on what Lyria trained on or what users may do with the output commercially.

What the API release added (refreshed 2026-10-01)

Google opened the model on 2026-09-04: Google AI Studio, the Gemini API and the Gemini app, beyond Flow Music. The model page (ai.google.dev/gemini-api/docs/models/lyria-3.5, last updated 2026-09-16) and the music-generation guide fill in some specs the announcement left out:

• Model id lyria-3.5, marked stable. Text and image input (up to 10 images), MP3 audio plus lyrics text out, 44.1 kHz stereo.
• Length: "a couple of minutes", steered by the prompt. No hard limit is given. A sibling, lyria-3-clip-preview, always returns 30 seconds.
• Single-turn: a generated track cannot be edited across prompts through the API. Flow Music's section editing is a product feature, not an API one.
• Filters: prompts asking for a specific artist's voice or copyrighted lyrics are blocked.

Still not published anywhere we found: training data, pricing on the model page, and any benchmark or listening test. So the T3 grade below still holds for the performance claims; the specs above come from Google's own developer docs.

Why it lands here anyway

A vocal song generator from a frontier lab is a structural event for the branch even without numbers: music generation had been a specialist market (suno, udio, ElevenLabs Music) while the big labs shipped speech (gemini-live-3-5-translate) and open instrumental research (musicgen, stable-audio). The closed frontier now competes directly on vocals + lyrics, the exact capability the open wedge does not have.

Tier T3, and unusually thin even for a vendor post — four adjectival bullets and a demo link, with nothing higher available at announcement time. Treat every claim here as a marketing assertion pending independent measurement.

Related

audio-music-generation · suno · udio · stable-audio · ai-music-copyright · audio-deepfake · gemini-live-3-5-translate · google (llm-providers-wiki)