Lyria 3.5 – Google DeepMind's AI Music Generation Model

1. What is Lyria 3.5

Lyria 3.5 is a new generation AI music generation model launched by Google DeepMind in July 2026, directly integrated into the Google Flow Music platform. This model achieves four key upgrades in terms of melodic structural complexity, lyric generation quality, vocal expressiveness, and creative control capabilities, moving AI music generation from an experimental tool to a practical creative assistant. By parameterizing the control of rhythm and duration, and trained on authorized music data, Lyria 3.5 strikes a balance between musicality and compliance, offering content creators an efficient, legal, and high-quality music generation solution.

Image source: Official article
Image source: official article

Technical positioning and domain: Lyria 3.5 belongs to the AI music generation domain, focusing on generating complete songs or music segments from text prompts, covering melody, lyrics, vocals, and orchestration. Unlike general-purpose audio generation models, Lyria 3.5 specifically addresses issues of structural coherence, emotional expression, and controllability in music creation, positioning itself as a professional creative assistant for content creators.

Development background: This model was developed by Google DeepMind, whose team has deep expertise in the field of audio generation (such as WaveNet, AudioLM, and the Lyria series). Lyria 3.5 represents the third major update from DeepMind within five months, reflecting its ongoing commitment to the music generation track. The model is trained using authorized music data to avoid copyright risks and includes built-in SynthID watermark technology for content traceability.

Core value: Lyria 3.5 addresses common pain points in previous AI music generation systems, such as a "mechanical" feel, disconnection between lyrics and prompts, lack of emotional expression in vocals, and insufficient creative control. It enables users to quickly generate music drafts of professional quality, significantly reducing the time and financial costs of music production while ensuring the legality and usability of the generated content.

Technical features: The model is based on the Transformer architecture, combining autoregressive and diffusion techniques to achieve high-fidelity audio generation. Its differentiating capabilities include: complex and natural transitions between melody and harmony, structured adherence of lyrics to prompts, emotional prosody modeling for vocals, and intuitive control over rhythm/duration parameters. Additionally, Lyria 3.5 is fully embedded within the Google Flow Music platform, requiring no local deployment and offering out-of-the-box usability.

2. Key Features

• Melody Generation: The model can create richly structured and naturally flowing melodies, significantly reducing the mechanical repetition often found in traditional AI-generated music. Through deep learning and harmonic progression modeling, it supports a wide range of styles—from classical to electronic—with melodies that follow stronger musical logic.

• Lyrics Creation: The generated lyrics excel in thematic consistency and structural integrity of paragraphs, with a greatly improved adherence to user prompts. Users can specify advanced attributes such as rhyme patterns and the number of sections, making the output more aligned with their creative intent.

• Vocal Synthesis: The output vocals feature clear articulation and emotional variation, closely resembling real human singing. By employing advanced acoustic models and prosody prediction techniques, the synthesized vocals achieve more nuanced expression in timbre, dynamics, and emotional delivery, significantly enhancing the emotional impact of the songs.

• Creative Control: The model allows direct adjustment of the output music's tempo (BPM) and total duration, meeting precise requirements across various scenarios such as short videos and podcasts. Parameterized control makes music generation more accurate, facilitating synchronization with video visuals or program pacing.

• End-to-End Workflow: Within the Google Flow Music platform, the process from inputting prompts to exporting finished audio is seamless, without the need to switch to external tools. The platform includes built-in post-production features such as mixing and mastering, streamlining the creative process and reducing technical barriers.

• Real-Time Preview and Iteration: Instant audio previews are available during the generation process, allowing users to adjust prompts or parameters and regenerate at any time. This interactive iterative mode greatly improves creative efficiency, enabling rapid trial and error until the desired result is achieved.

• Copyright Compliance and Traceability: All training data comes from authorized sources, and the generated content includes an embedded SynthID digital watermark for easy identification of AI-generated attributes. This provides legal protection for commercial use and helps avoid copyright disputes.

3. How to Use

1. Access and Login: Open the official website of Google Flow Music (flowmusic.app) and log in with a Google account. It is recommended to use the latest versions of Chrome, Edge, or Safari browsers to ensure proper functionality of audio processing and real-time preview features.

2. Create a New Project: Click "New Song" in the creation interface to enter the prompt input panel. You can first set macro parameters, such as music style (pop, rock, electronic, etc.), emotional tone (upbeat, melancholic, intense, etc.), and reference artists or songs.

3. Enter Prompts: Describe the desired musical content and structure in natural language. For example: "A lively pop electronic track with lyrics about a summer beach, including verses and choruses." The more specific the prompt, the more aligned the generated results will be with your expectations.

4. Adjust Creation Controls: Set the target tempo (BPM, range 60–200) and duration (15 seconds to a full song, up to approximately 5 minutes) in the parameter panel. You can also choose whether to generate vocals, select vocal types (male, female, chorus), and choose the language of the lyrics (currently primarily supporting English).

5. Generate and Preview: Click the "Generate" button to start the model's creation process. The first version of the music snippet is typically output within a few seconds, with real-time playback preview available. If you are not satisfied, you can immediately modify the prompt or parameters and regenerate.

6. Fine-Tuning: Use Flow Music's editing tools to make fine adjustments to the generated output, including track volume balancing, adding audio effects (reverb, compression, etc.), and replacing specific sections. Manual editing of the lyrics text is supported, followed by re-synthesis of the vocals.

7. Export the Final Product: After confirming the final version, select the export format (WAV or MP3) and audio quality (sampling rate, bit rate), then download it to your local device. The exported audio files can be used directly for video soundtracks, podcasts, or commercial projects.

8. Notes: Generated content is subject to Google's service terms and is intended for legal use only. It is recommended to verify the watermark using the SynthID detection tool before public use to indicate AI-generated attributes. Additionally, generating high-complexity songs may require waiting 30–60 seconds; please ensure a stable internet connection.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For commercial users requiring high-quality, controllable, and copyright-safe tools (such as advertising agencies and video creators), Lyria 3.5 is the most reliable option. Its licensed training data and built-in watermark mechanism significantly reduce legal risks, while Flow Music's end-to-end workflow simplifies the creative process. However, note its platform lock-in and language limitations; if the project involves multilingual lyrics or requires offline deployment, careful evaluation is necessary.

For developers seeking fast generation of multi-style music, emphasizing community sharing and API integration, Suno v4 offers more flexible access methods and a rich style library. However, its copyright controversies in training data may pose risks for commercial use, and it is recommended for non-commercial scenarios first. For professional music producers, Udio excels in audio quality and instrument separation, but its control capabilities are limited, making it more suitable as an inspiration tool rather than for final output. The open-source nature of Meta MusicGen makes it the preferred choice for researchers and those with customized needs, but its music quality and functional completeness lag behind commercial models, making it more appropriate for technical validation and academic exploration.

6. Editor's Summary

Lyria 3.5 has achieved a significant leap in the AI music generation field, moving from "listenable" to "pleasurable." The improvements in melodic complexity and vocal expressiveness are attributed to DeepMind's years of technical accumulation in audio generation, particularly in modeling long-range dependencies in musical structures and achieving fine-grained control over emotional rhythms. Compared to competitors, the biggest differentiating advantage of Lyria 3.5 lies in its copyright compliance — all training data is authorized, and it integrates SynthID watermarking, which holds significant commercial value in the current environment of frequent AI music copyright disputes. Additionally, its deep integration with Google Flow Music reduces the barrier to entry, allowing non-professional users to quickly generate usable music.

In terms of practical value, Lyria 3.5 provides content creators such as short video producers, advertisers, and podcasters with an efficient, low-cost music generation solution, especially suitable for scenarios requiring rapid iteration and precise control over duration. Its parameterized rhythm adjustment feature simplifies and directly enables synchronization between background music and visuals. However, the model currently only supports English lyrics, and its level of control is not yet on par with professional music production software, which limits its application in high-end music creation.

The target users mainly include: video creators, podcasters, and advertising professionals who need to quickly obtain original music; music enthusiasts looking to aid in creative inspiration; and independent game and film developers requiring a large number of demo soundtracks. For professional music producers, Lyria 3.5 is more suitable as a draft tool rather than a final product.

In terms of future development potential, the architecture of Lyria 3.5 leaves room for further iterations. With DeepMind's continued investment in music generation, future versions are expected to expand multilingual support, add more detailed control over instrument arrangements, and possibly open API interfaces to integrate into a broader creative ecosystem. Meanwhile, collaboration with multimodal models such as Gemini could enable cross-modal creation from text, images, to music, further broadening its application boundaries.

7. Application Scenarios

• Short Video Music Generation: Rapidly generate rhythm-matched and duration-precise background music for platforms such as Douyin, YouTube Shorts, and Instagram Reels. Users simply need to input style and emotion keywords, adjust BPM and duration, and obtain original music synchronized with the video, avoiding copyright disputes.

• Podcasts and Audiobooks: Customize intro/outro music and chapter transition sound effects to enhance content professionalism. The lyric generation feature in Lyria 3.5 can also be used to create podcast theme songs or sponsor jingles, improving brand recognition through voice synthesis.

• Advertising and Marketing: Generate original ad jingles or background melodies that align with brand tone, avoiding legal risks associated with unauthorized music use. Marketing teams can quickly produce multiple versions for A/B testing, optimizing ad performance.

• Game and Film Prototyping: Quickly generate demo music for indie games and animated short films, accelerating early-stage iteration. The end-to-end workflow of Lyria 3.5 allows directors or producers to obtain audio references close to the final result during the creative phase, without waiting for professional composers.

• Personal Music Creation: Assist musicians in ideation, melody drafting, or full song prototyping. Users can first generate a basic framework using Lyria 3.5, then import it into a DAW for detailed editing, significantly shortening the time from inspiration to final product.

8. FAQ

Q: Is Lyria 3.5 free to use?
A: Google Flow Music currently offers a free trial quota, but full functionality may require a subscription to Google One or purchasing a standalone plan. Specific pricing details will be determined by Google's official announcement. We recommend visiting the Flow Music official website to check the latest plans.

Q: Who owns the copyright to the generated music?
A: According to Google's service terms, users have the right to use content generated via Lyria 3.5, but must comply with relevant laws and regulations. Since all training data comes from authorized sources, users can use the generated music in commercial projects without additional copyright fees. However, it is recommended to verify whether the content includes third-party elements before publishing.

Q: Does it support generating Chinese lyrics?
A: Currently, Lyria 3.5 is primarily optimized for English, and the quality of Chinese lyric generation is limited. There may be grammatical errors or unnatural pronunciation issues. If you need Chinese lyrics, it is recommended to first generate them in English and then manually translate, or wait for future version updates.

Q: Can I export stem files?
A: The Flow Music platform supports basic audio editing and mixing, but stem export functionality is not yet publicly available. At present, only full mixed audio files (WAV/MP3) can be exported. If you need stem processing, you can import the final product into a DAW for further separation.

Q: How can I ensure that the generated content does not infringe on rights?
A: All training data for Lyria 3.5 comes from authorized sources, and the generated content includes a built-in SynthID digital watermark, which helps the platform and users identify AI-generated attributes. Users should avoid directly referencing copyrighted song titles or artist styles in prompts to prevent potential legal risks.

Q: How long does it take to generate a complete song?
A: Typically, generating a 30-second music clip takes 5–10 seconds, while a full 3-minute song (including vocals) may take 30–60 seconds. Generation speed is affected by network conditions, song complexity, and current server load.

9. Project Links

• Lyria 3.5 Official Model Page: https://deepmind.google/models/lyria/
• Google Flow Music Official Website: https://www.flowmusic.app/
• Google DeepMind Official Organization: https://deepmind.google/
• SynthID Watermark Technology Introduction: https://deepmind.google/technologies/synthid/