Lyria 3.5

Lyria 3.5 is a generative AI model developed by Google DeepMind specifically for high-fidelity music creation. It is designed to generate full-length songs with complex structural coherence, including distinct sections such as verses, choruses, and bridges. The model produces high-quality 44.1 kHz stereo audio in MP3 or WAV formats, offering improvements in musicality and vocal expressiveness over previous iterations in the Lyria family.

The model is multimodal, allowing users to generate music from either text descriptions or image inputs. It can process up to 10 reference images to inform the mood, style, or atmosphere of a track. Lyria 3.5 supports variable song lengths, ranging from 30-second clips to full compositions lasting up to three minutes. It also features refined vocal generation across multiple languages, including English, Spanish, French, German, and Japanese.

To ensure responsible use and content provenance, all audio generated by Lyria 3.5 includes SynthID watermarking. This technology embeds an inaudible digital watermark directly into the audio waveform, which remains detectable even after common edits like compression or speed adjustments. The model is also integrated with safety guardrails designed to prevent the unauthorized imitation of specific recording artists.

For optimal results, users can employ structural tags like [Verse], [Chorus], and [Bridge] within their text prompts to guide the arrangement. Providing specific details regarding genre, instrumentation, tempo (BPM), and mood typically enhances the model's adherence to the desired output. Developers can access the model via the Gemini API, where it supports fine-grained duration and structural controls.