Lyria

What it is

Lyria is a music-generation model from Google DeepMind, announced on 16 November 2023^[1]. Where a language model produces prose and an image model produces pictures, Lyria produces audio: Google DeepMind said it excels at generating high-quality music with instrumentals and vocals, and that it can perform continuation tasks — taking music that already exists and carrying it onward^[1].

Where it came from

The model did not arrive as a standalone product. It debuted in Dream Track, a trial inside YouTube Shorts that creates a 30-second soundtrack with an AI-generated voice in the style of participating artists. It also appeared in a set of Music AI tools designed with artists and producers^[1]. Both settings put the model in the hands of a defined group of users rather than the open public, and both tied the technology to people who make music for a living.

From the start, Google attached a provenance marker to the output: Lyria's audio carries an inaudible SynthID watermark, designed to survive common modifications such as MP3 compression^[1]. Generated music is easy to copy and re-encode, so a mark that persists through compression gives a way to test whether a track came from the model.

Later versions

Lyria 2 was announced at Google I/O in May 2025. It was aimed at creators through YouTube Shorts and at enterprises on Vertex AI, and a real-time variant, Lyria RealTime, became available through the API and AI Studio^[2]. A preview of the real-time model, lyria-realtime-exp, was released in the Gemini API on 20 May 2025, with no shutdown date announced^[3].

Lyria 3 previews arrived in the Gemini API on 25 March 2026, under the identifiers lyria-3-pro-preview and lyria-3-clip-preview, and were followed by lyria-3.5 on 3 September 2026^[3]. Google calls Lyria 3 Pro Preview its flagship music generation model^[4]. That version is optimized for full-length songs with multiple verses, choruses and bridges, and generates high-quality 48kHz stereo audio from text prompts or image inputs^[4].

How the pieces fit together

The Lyria family has grown along three separate routes at once. One is consumer-facing: short soundtracks inside YouTube Shorts, first as the Dream Track trial and later as one of the surfaces for Lyria 2^[1][2]. The second is enterprise and developer access: Lyria 2 for enterprises on Vertex AI, and the Lyria 3 previews in the Gemini API^[2][3]. The third is real-time generation, where the model produces audio as it is being used rather than in a single pass — Lyria RealTime through the API and AI Studio, and the lyria-realtime-exp preview^[2][3].

The naming shows how the releases were staged. Previews with explicit identifiers sit alongside flagships, and Google's developer documentation records when each preview appeared; for lyria-realtime-exp it also notes that no shutdown date had been announced^[3]. Earlier previews were followed by newer versions rather than staying fixed, which is a useful reminder that a model name in an API is a moving target.

What has stayed constant across the versions is the combination Google emphasizes: instrumental and vocal music together, control over structure — continuation at the beginning, full-length songs with multiple verses, choruses and bridges at the flagship end — and a watermark carried inside the audio itself^[1][4]. The 48kHz stereo output and the option of image inputs on Lyria 3 Pro Preview extend the same idea: give the model something to work from, and let it return finished audio rather than a sketch^[4].