Generative AI Google Brain

audio-music-getting-started-lyria2-music-generation

2 min read

Lyria 2 Music Generation

Source notebook

Repo path: audio/music/getting-started/lyria2_music_generation.ipynb · Open on GitHub · intro

Generates 30-second 48 kHz WAV music clips from text prompts with Lyria 2 on Vertex AI.

Summary

This notebook teaches how to call Google’s Lyria 2 music generation model through a Vertex AI prediction endpoint. It authenticates with Application Default Credentials, builds a REST request with prompts and parameters, decodes base64 WAV output, and plays generated audio in the notebook. The examples demonstrate genre, mood, tempo, instrumentation, negative prompts, sample counts, and seed-based deterministic generation.

Key code patterns

Authenticated REST call

creds, project = google.auth.default()
auth_req = google.auth.transport.requests.Request()
creds.refresh(auth_req)
headers = {
    "Authorization": f"Bearer {creds.token}",
    "Content-Type": "application/json",
}
response = requests.post(api_endpoint, headers=headers, json=data)
response.raise_for_status()

Shows how the notebook obtains an access token and calls the prediction endpoint directly.

Lyria predict endpoint

music_model = (
    f"https://us-central1-aiplatform.googleapis.com/v1/"
    f"projects/{PROJECT_ID}/locations/us-central1/"
    "publishers/google/models/lyria-002:predict"
)

Defines the regional Vertex AI endpoint for the Lyria 2 model.

Music generation request

req = {"instances": [request], "parameters": {}}
resp = send_request_to_google_api(music_model, req)
return resp["predictions"]

Wraps prompt settings into the predict request shape expected by the model.

Decode and play audio

bytes_b64 = dict(pred)["bytesBase64Encoded"]
decoded_audio_data = base64.b64decode(bytes_b64)
audio = Audio(decoded_audio_data, rate=48000, autoplay=False)
display(audio)

Converts model output from base64 into playable 48 kHz audio in the notebook.

Models & APIs used

  • Models: lyria-002
  • APIs / services: Vertex AI, Agent Platform API
  • SDKs / libraries: google.auth, requests, IPython.display

When to use this

Use this pattern when generating short music clips from text prompts with Lyria 2 through Vertex AI REST predictions.

Gotchas & caveats

  • Requires an existing Google Cloud project.
  • The Agent Platform API must be enabled through aiplatform.googleapis.com.
  • Colab users must authenticate with google.colab.auth.authenticate_user().
  • The endpoint is hard-coded to us-central1.
  • seed and sample_count cannot be set in the same request.
  • Generated clips are described as 30 second WAV audio at a 48 kHz sample rate.

Best practices

  • Use detailed prompts describing style, mood, tempo, rhythm, and instrumentation.
  • Use negative_prompt to specify audio qualities to exclude.
  • Use seed for deterministic generation when not using sample_count.
  • Decode bytesBase64Encoded output before playing the audio.
  • Rely on response.raise_for_status() to surface failed API calls.

Graph View

Backlinks