Noise2Music
Diffusion-based text-to-music: 30-second 24kHz clips from freeform prompts.
About
Noise2Music was one of the first diffusion models aimed squarely at text-to-music, published by Google Research in early 2023. It chains diffusion models in a cascade: a generator turns a text prompt into an intermediate representation, either a log-mel spectrogram or a low-fidelity 3.2kHz waveform, and further stages lift that to a 16kHz and then a 24kHz waveform. The result is a 30-second music clip that follows freeform descriptions of genre, instruments, mood and activity.
A useful byproduct was MuLaMCap, a 400K music-text pair dataset built by annotating AudioSet's music content with a large vocabulary of descriptive phrases. Evaluation used Frechet Audio Distance and MuLan text-audio similarity, both provider-reported. Noise2Music was never released publicly and has since been superseded inside Google by MusicLM and Lyria, but it remains a landmark in diffusion-based music generation.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
FAD on AudioSet-music and MagnaTagATune plus MuLan similarity (provider-reported); no independent replication found.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: Google Research paper, influential early diffusion music work, MuLaMCap dataset contribution. Strengths: cascaded diffusion design, fine-grained text control, solid dataset byproduct. Docked for: (1) never publicly released; (2) superseded by MusicLM and Lyria; (3) evals are provider-reported. License: proprietary research, no public release.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 1.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Google Research, arXiv 2302.03917, early 2023
- · Cascaded diffusion: text prompt to 30s 24kHz music
- · MuLaMCap: 400K music-text pair dataset byproduct
- · Superseded by MusicLM/Lyria; never publicly released
- · Google Research brand logo staged