Basic music generation

A short prompt is enough to get useful 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 runs against either Google Lyria (30 s MP3 clip; lyria-3-clip-preview) or ElevenLabs Music (full song up to 5 min; music_v1), depending on --provider=.

Scenario. You need background audio for a video, a song draft for a demo, or a quick musical sketch. You want to compare what two providers do with the same prompt.

The recipe body is provider-agnostic

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

const result =
  yield *
  generate({
    model: "music_v1", // or "lyria-3-clip-preview"
    prompt: "Lo-fi piano with brushed drums, 70 BPM, melancholic",
    duration: Duration.seconds(30),
    outputFormat: { container: "mp3", encoding: "mp3", sampleRate: 44100, channels: 2 },
  })

yield * writeFile("out.mp3", result.primary.audio.bytes)

The recipe yields the generic MusicGenerator service. The runner (run.ts) picks the Layer based on --provider=, so the same body works against both providers without changes.

Run it

# Default: ElevenLabs with the built-in prompt
ELEVENLABS_API_KEY=... pnpm tsx recipes/basic-music-generation/run.ts

# Explicit provider
GOOGLE_API_KEY=...     pnpm tsx recipes/basic-music-generation/run.ts --provider google
ELEVENLABS_API_KEY=... pnpm tsx recipes/basic-music-generation/run.ts --provider elevenlabs

# Custom prompt from a .txt file, 75-second clip
ELEVENLABS_API_KEY=... pnpm tsx recipes/basic-music-generation/run.ts \
  --provider elevenlabs --duration 75 --prompt-file ./my-prompt.txt

Flags (--name value and --name=value both work):

Flag	Default	Notes
`--provider`	`elevenlabs`	`elevenlabs` or `google`.
`--prompt-file`	built-in	Path to a `.txt` file. A bare positional path is not picked up.
`--duration`	`30`	Clip length in seconds. ElevenLabs honors it; Lyria `clip` ignores it.

Audio lands in output/basic-music-generation/<timestamp>/track.mp3.

Writing a prompt that gets followed

In prompt mode ElevenLabs treats the text as a style brief. Things that made a real difference while iterating on this recipe:

Spell the lyrics out. Describing the lyrics (“a song that celebrates…”) yields improvised, usually English, lyrics. Put the actual lines in the prompt under section tags ([Verse 1], [Chorus], [Bridge]) and say “with these exact lyrics”.
Give it room. 30 s is too short for a verse and a chorus with names in it. 60–120 s works for a short song.
Name the accent and pronunciation. State the singer’s origin and a couple of concrete phonetic cues (e.g. “castilian z, no seseo”). Spell out how unusual names should sound.
Never name real artists. “In the style of ” is rejected with a 4xx InvalidRequest. Describe the genre instead.

Where the providers differ

The shared MusicGenerator surface intentionally hides most differences. The ones worth knowing:

Capability	Lyria 3 sync	ElevenLabs Music
Duration	Fixed 30 s for `clip`, controllable on `pro`.	Honored. `music_length_ms` 3 s – 10 min.
Streaming	Single chunk after sync (fake stream).	Native chunked HTTP via `POST /v1/music/stream`.
Lyrics	No structured wire field; embed in prompt.	Per-section `lines` in composition plan, or embed in prompt.
Watermark	Always `"synthid"` (mandatory).	Opt-in `"c2pa"` via `signWithC2pa: true` (MP3 only).
Vocals control	Embed `"no vocals"` in your prompt.	`forceInstrumental` on the typed request.
Composition plan (per-section lyrics + styles + durations)	.	Yes (`compositionPlan`). Free plan-generator at `POST /v1/music/plan`.

These provider-specific knobs live on the typed services (LyriaGenerator, ElevenLabsMusicGenerator), not on the cross-provider CommonGenerateMusicRequest. Reach for them when the shared surface isn’t expressive enough.

What this generalises to

The recipe body returns Effects that yield the generic MusicGenerator service. Future providers (Suno, Mureka, MiniMax, Stable Audio, Tencent SongGen) can hide their job polling, streaming, or session setup behind the same calls. The music generation overview covers the cross-provider shapes and the capability marker that gates bidirectional sessions.

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