Lyria API

Lyria is a new foundation model for high-quality audio generation, capable of creating diverse soundscapes and musical pieces from text prompts. Lyria enables users to generate high-quality instrumental music from text prompts.

To explore this model in the console, see the Lyria model card in the Model Garden (accessible using the Media Studio tab).

Try Lyria on Vertex AI (Vertex AI Studio)

Try Lyria in a Colab

Supported Models

The Lyria API supports the following model:

  • lyria-002

HTTP request

curl -X POST \
  -H "Authorization: Bearer $(gcloud auth print-access-token)" \
  -H "Content-Type: application/json" \
  https://LOCATION[-aiplatform.googleapis.com/v1/projects/](https://-aiplatform.googleapis.com/v1/projects/)PROJECT_ID/locations/LOCATION/publishers/google/models/lyria-002:predict \
  -d '{
    "instances": [
      {
        "prompt": "string",
        "negative_prompt": "string", // Optional
        "seed": 0 // Optional. Cannot be used with sample_count.
      }
    ],
    "parameters": {
      "sample_count": 1 // Optional. Cannot be used with seed.
    }
  }'

Use the following parameters for the Lyria model. For more information, see the Lyria Model Garden card details.

Parameter

prompt

(in instances object)

string

Required. The text description in US English (en-us) of the audio to generate.

Example: "An energetic electronic dance track with a fast tempo."

negative_prompt

(in instances object)

string

Optional. A description of what to exclude from the generated audio.

Example: "vocals, slow tempo"

seed

(in instances object)

integer

Optional. A seed for deterministic generation. If provided, the model will attempt to produce the same audio given the same prompt and other parameters.

Cannot be used with sample_count in the same request.

Example: 12345

sample_count

(in parameters object)

integer

Optional. The number of audio samples to generate. Default is 1 if not specified and seed is not used.

Cannot be used with seed in the same request.

Example: 2

Sample request

Use the following request to generate instrumental music from a text prompt:

Text-to-music generation request

curl

curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
https://us-central1-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/us-central1/publishers/google/models/lyria-002:predict \
-d '{
  "instances": [
    {
      "prompt": "A calm acoustic folk song with a gentle guitar melody and soft strings.",
      "negative_prompt": "drums, electric guitar",
      "seed": 98765
    }
  ],
  "parameters": {}
}'

JSON

This example uses seed for reproducible output.


{
  "instances": [
    {
      "prompt": "A calm acoustic folk song with a gentle guitar melody and soft strings.",
      "negative_prompt": "drums, electric guitar",
      "seed": 98765
    }
  ],
  "parameters": {}
}

JSON with sample_count

This example uses sample_count to generate multiple samples.

To test a text prompt using the Lyria API, send a POST request to the publisher model endpoint. The following example omits seed from the instances object and uses sample_count in the parameters object to generate multiple samples. 


{
  "instances": [
    {
      "prompt": "A calm acoustic folk song with a gentle guitar melody and soft strings.",
      "negative_prompt": "drums, electric guitar"
    }
  ],
  "parameters": {
    "sample_count": 2
  }
}

Response body

A successful request returns a JSON object containing the generated audio data. Each generated audio clip is 30 seconds long and provided as a WAV audio file at a 48kHz sample rate.

{
  "predictions": [
    {
      "audioContent": "BASE64_ENCODED_WAV_STRING_SAMPLE_1",
      "mimeType": "audio/wav"
    }
    // Additional audio samples will be listed here if sample_count > 1
    // e.g.,
    //{"audioContent": "BASE64_ENCODED_WAV_STRING_SAMPLE_2",
    // "mimeType": "audio/wav"
    //}
  ],
  "deployedModelId": "xxxxxxxxxxxxxxx", // Actual ID may vary based on deployment
  "model": "projects/PROJECT_ID/locations/LOCATION/publishers/google/models/lyria-002",
  "modelDisplayName": "Lyria 2"
}
Response element
predictions

array

An array of generated audio samples. Each object in the array represents one audio clip.
predictions[].audioContent

string

Base64-encoded string of the generated WAV audio data.
predictions[].mimeType

string

The MIME type of the audio data. For Lyria, this is "audio/wav".
deployedModelId

string

The ID of the deployed model that processed the request (if applicable for the endpoint type).
model

string

The full resource name of the model that processed the request.
modelDisplayName

string

The display name of the model.

Best practices and limitations

Refer to the Lyria Model Card for detailed best practices on prompting, language support (US English only for prompts), generation times, output format (WAV, 48 kHz, 30s instrumental clips), safety measures, and deployment information.

Key points:

  • Detailed Prompts: Generally lead to better audio.
  • Specify: Genre, mood, instrumentation, tempo.
  • Negative Prompting: Use negative_prompt to exclude elements.
  • Output: 30-second WAV audio clips, 48 kHz, instrumental only.
  • Safety: Content safety filters, recitation checking, artist intent checks, and SynthID watermarking are applied.

Pricing

Lyria 2 usage is priced at $0.06 per 30 seconds of output music generated. For more details, see Vertex AI Pricing.

More information

  • Learn more about Generative AI on Vertex AI.
  • For an overview of Lyria, refer to its model card available in the Model Garden (Media Studio).

What's next