Google's Lyria 3.5 doubles as a text-to-speech model with a clever prompt trick
arnicas · x · 2026-09-07
- fofrAI discovered Google's music model Lyria 3.5 can be prompted as a text-to-speech model: asking for e.g. a "jaded, candid and natural" British millennial woman talking, with "no lyrics, no music, no beats," for 60 seconds yields surprisingly realistic raw speech clips.
- arnicas noted surprise that lyrics can carry speech and background sounds, and that mixing them with music still works—calling it genuinely useful.
Related event: Lyria 3.5 Can Work as TTS, Generating Realistic Speech from One Prompt(2 posts)→
More from Multimodal
- Small open-source TTS is getting interesting: Hojo-TTS-Light ships 40M and 80M variants — ZabihullahAtal · 2026-10-23
- Obsidian Light Mirage prompt template sculpts dramatic chiaroscuro portraits — LudovicCreator · 2026-10-23
- AI-generated fantasy short 'The Sky Doesn't Kneel' stuns with dragon-rider visuals — scenemixer_official · 2026-10-23
- Stunning AI-generated Dark Knight images nail foggy Batman atmosphere — egeberkina · 2026-10-23
- PixText: a sub-1B model that fixes mangled text in AI-generated images — Creepy_Astronomer_83 · 2026-10-23
- AI video now pulls off orbiting, diving camera moves around a speeding bicycle — umesh_ai · 2026-10-23