Basic music generation

A short prompt is enough to get useful background music.

Give the model a vibe, a tempo hint, or a rough song structure and it returns audio bytes you can write to disk. The recipe shows the small path first, then a richer prompt with weighted influences and lyrics.

Scenario. You need background audio for a video, a song draft for a demo, or a quick musical sketch. You want a 30-second clip from a prompt, not a local model or a new provider SDK.

Simple Or Directed

import { generate } from "@effect-uai/core/MusicGenerator"

// Simple
const simple =
  yield *
  generate({
    model: "lyria-3-clip-preview",
    prompts: "Lo-fi piano with brushed drums, 70 bpm",
  })
yield * writeFile("out-simple.mp3", simple.bytes)

// Weighted
const weighted =
  yield *
  generate({
    model: "lyria-3-clip-preview",
    prompts: [
      { text: "minimal techno", weight: 0.7 },
      { text: "ambient pad", weight: 0.3 },
    ],
    bpm: 124,
    scale: "A_MINOR",
    lyrics: "[Verse]\nA late train hums beneath the city\n",
  })
yield * writeFile("out-weighted.mp3", weighted.bytes)

Both calls use the same MusicGenerator.generate boundary: prompt in, MusicResult out. The simple prompt is enough for most demos. The weighted prompt is useful when you want to blend influences, preserve a song section shape, or carry musical hints like bpm and scale.

lyria-3-clip-preview returns a fixed 30-second MP3. Use lyria-3-pro-preview in index.ts when you want longer clips or WAV output.

Lyria adds a SynthID watermark. The adapter surfaces that as result.watermark, so downstream code can preserve provenance.

Run it

# Both built-in variants
GOOGLE_API_KEY=... pnpm tsx recipes/basic-music-generation/run-node.ts

# Custom simple prompt from a .txt file
GOOGLE_API_KEY=... pnpm tsx recipes/basic-music-generation/run-node.ts ./my-prompt.txt

# Custom weighted prompt from a .json file (see index.ts for the WeightedConfig shape)
GOOGLE_API_KEY=... pnpm tsx recipes/basic-music-generation/run-node.ts ./my-track.json

Writes out-simple.mp3 and/or out-weighted.mp3 next to the recipe.

What This Generalizes To

Today this runs on Google’s Lyria layer, but the recipe body yields the generic MusicGenerator service. Future providers can hide job polling, streaming, or session setup behind the same boundary.

For the broader model surface and planned interactive sessions, see Music generation. For spoken audio instead of music, see Basic speech synthesis.

The full source lives next to this README at index.ts.