Google rolls out MusicFX for text-to-track generation

Turning a vague mood into an actual piece of music used to require an audio workstation or hours of digging through royalty-free libraries. Now Google is making that process as simple as typing a sentence.

Google officially opened up two generative AI audio experiments in its AI Test Kitchen: MusicFX, a text-to-music generator, and TextFX, a creative writing assistant designed for lyricists. Both tools had been floating around in private previews since December, but they are now broadly accessible.

Prompt to loop in 70 seconds

MusicFX builds on top of MusicLM, the text-to-audio model Google showed off last year. The search giant claims this updated version is both faster and delivers higher fidelity.

Type in a descriptive phrase, like two nylon string guitars playing in flamenco style, and the system serves up two 30-second clips. If you need more breathing room, you can extend those snippets to 50 or 70 seconds, or hit a button to automatically loop the start and end into an unbroken beat.

Google also built in prompt suggestions to guide users. If you type country style, a drop-down suggests rockabilly or bluegrass. Type catchy, and it offers chill or melodic. Below the prompt box, a floating word cloud surfaces instruments, tempos, and vibes, from 808 drums to avant-garde.

In practice, the tracks sound crisp and clean, but somewhat safe and uninspired. A danceable house groove with funky beats comes out reasonably pleasant, but anything leaning on orchestral strings quickly turns brittle, sounding more like cheap MIDI samples than an actual scoring session.

Rap-assisted lyric blocks and copyright tripwires

To avoid getting dragged into court by music publishers, Google wrapped MusicFX in heavy guardrails. Prompts requesting vocal performances or naming specific recording artists are flatly rejected. Every audio file also carries SynthID, an imperceptible digital watermark developed by DeepMind to prove the track originated from Google AI.

That blocklist has some holes. While it refuses prompts styled after SZA or The Beatles, it had no problem generating tracks referencing indie outfits like Lake Street Dive.

Alongside music generation, Google expanded TextFX, a lyric companion tool developed with rapper and producer Lupe Fiasco. Powered by PaLM 2, TextFX rejects the idea of hitting a button to get a generic verse. Instead, it offers a suite of playful language modules. One module helps writers surface alliterative words within a category, while another finds surprising metaphors connecting two unrelated concepts.

It takes some getting used to, but it feels genuinely practical for lyricists and copywriters hunting for unexpected rhymes. Keep an eye on its factual accuracy, though. Google warns that TextFX can spit out factual errors, and in hands-on testing, it even parroted debunked conspiracy theories.

Why it matters

For creators, MusicFX is a convenient sketchpad. If you need a fast, non-copyrighted 30-second loop for a podcast intro, a mobile game prototype, or a short-form video background, it gets the job done at zero cost. But without vocal generation and with synthetic strings still sounding flat, it is not replacing studio producers anytime soon.

Across the industry, Google is walking a fine line. While companies like Stability AI lean on fair-use arguments to train on scraped material, and record labels aggressively scrub unauthorized voice clones off Spotify, Google is trying to play nice. Between its artist-backed YouTube AI experiments and built-in SynthID watermarking, the company is positioning itself as the responsible adult in generative audio.

We think this ultra-cautious approach is smart legal armor against the recording industry, but it comes at a creative cost. Stripping out vocals and playing it overly safe keeps Google out of the courtroom, but it also leaves MusicFX feeling more like a polished corporate demo than an inspiring creative powerhouse.