Lyria

Lyria - AI Music Generation Skill

Use the Python script in scripts/ to generate music via Google Gemini Lyria 3. The model is automatically selected based on the request / purpose:

Lyria 3 Pro is auto-selected when any of these conditions are met (OR):

• --pro flag is specified
• Prompt contains full-song keywords (e.g. 歌詞, フル尺, 楽曲, full-length, verse, chorus, lyrics, professional)
• Prompt is 100+ characters long

Lyria 3 Clip is selected when:

• --clip flag is specified
• Prompt contains clip keywords (e.g. ループ, ジングル, 30秒, clip, loop, jingle, preview)
• None of the Pro conditions are met (Clip is the default — fast and cheap, ideal for iterating on prompts)

Tip: Iterate on your prompt with the faster --clip model first, then commit to a full-length generation with --pro.

Prerequisites

1. Install dependencies

pip install google-genai

2. Configure API credentials (one of the following)

Option A: Gemini Developer API (recommended for personal use)

Set the GEMINI_API_KEY environment variable. Get a key at https://aistudio.google.com/apikey

export GEMINI_API_KEY="your-api-key"

Option B: Vertex AI API (for Google Cloud users)

Set GOOGLE_CLOUD_PROJECT and optionally GOOGLE_CLOUD_LOCATION. Requires a GCP project with the Vertex AI API enabled and Application Default Credentials configured (gcloud auth application-default login).

export GOOGLE_CLOUD_PROJECT="your-project-id"
export GOOGLE_CLOUD_LOCATION="us-central1"   # optional, defaults to us-central1

Priority: If both GOOGLE_CLOUD_PROJECT and GEMINI_API_KEY are set, Vertex AI is used.

3. Optional environment variables

Script

scripts/generate.py - Music generation

Basic generation (auto-selects model)

python scripts/generate.py "lofi hip hop, mellow piano, rainy night, relaxed beat"

Force a short clip (fast iteration / loops / jingles)

python scripts/generate.py "upbeat synthwave jingle, 80s, short loop" --clip -o jingle.mp3

Force a full-length song (with structure tags and lyrics)

python scripts/generate.py \
  "[Verse] city lights fade into the night [Chorus] we are still alive, dreaming, full pop ballad, female vocals" \
  --pro -o song.mp3

Request WAV output (Pro only; Clip outputs MP3)

python scripts/generate.py "epic orchestral cinematic, full-length, dramatic" --pro --wav -o cinematic.wav

Disable SSL verification (for corporate proxies or self-signed certs)

python scripts/generate.py "ambient pad, calm" --no-ssl-verify -o ambient.mp3

JSON output (for programmatic use)

python scripts/generate.py "jazz trio, walking bass" --json -o jazz.mp3

Full options

usage: generate.py [-h] [-o OUTPUT] [--pro] [--clip] [--wav] [-v]
                   [--json] [--no-ssl-verify] prompt

Arguments:
  prompt              Text prompt describing the music

Options:
  -o, --output PATH   Output file path (auto-generated if omitted; extension
                      derived from the returned audio MIME type)
  --pro               Force Lyria 3 Pro model (full-length songs)
  --clip              Force Lyria 3 Clip model (30s clips, loops, jingles)
  --wav               Request WAV output (Pro only; Clip outputs MP3)
  -v, --verbose       Show detailed output
  --json              Output result as JSON
  --no-ssl-verify     Disable SSL certificate verification

--pro and --clip are mutually exclusive.

Output Format

The output file extension is derived from the audio MIME type returned by the API.

Prompting Tips

A good music prompt combines several dimensions:

1. Genre / style — lofi hip hop, synthwave, orchestral, jazz, EDM
2. Mood / atmosphere — melancholic, uplifting, tense, dreamy
3. Instruments / arrangement — mellow piano, driving bass, string section
4. Tempo / rhythm — slow 70 BPM, fast four-on-the-floor, swing
5. Vocals (Pro) — female vocals, warm timbre, male rap, or instrumental

Song structure (Pro): use section tags so the model builds verses, choruses, and bridges:

[Verse] ... lyrics ...
[Chorus] ... lyrics ...
[Bridge] ... lyrics ...

You can also add timestamps for precise timing, e.g. [0:00 - 0:10] soft intro. The prompt language determines the sung language.

See references/prompts.md for a comprehensive prompt reference with category-specific templates.

Limitations

• Single-turn: each call is independent; there is no multi-turn editing of a previously generated track.
• No artist voice cloning: prompts that request a specific real artist's voice are blocked by safety filters.
• No negative prompts / seed: describe what you want directly in the prompt.

Error Handling