OpenL3
Open deep audio (and image) embeddings from self-supervised audio-visual correspondence.
About
OpenL3 is the open-source answer to DeepMind's Look, Listen and Learn: it trains an embedding by predicting whether an audio clip and a video frame come from the same video, then packages the result as a pip-installable library (pip install openl3) that spits out deep audio embeddings with no labels required. The ICASSP 2019 paper reports it outperforming VGGish, SoundNet, and the original L3-Net on several sound-recognition tasks.
For music work it is a solid general-purpose descriptor: people routinely use OpenL3 embeddings as features for instrument and genre classifiers, and the library's simplicity made it a staple baseline. It is MIT licensed and got TensorFlow 2 support after launch. The honest downside is age: the project has seen little development in recent years, the training code had a documented temporal-overlap bug (which the authors disclosed), and newer self-supervised models have passed it on raw accuracy. Still perfectly usable as a lightweight, well-understood embedding.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported: ICASSP 2019 paper reports beating VGGish, SoundNet, and L3-Net on sound-recognition tasks; provider-claimed, era-level results.
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, metrics-v1). Evidence: ICASSP 2019 paper (Cramer et al., NYU MARL); MIT-licensed repo marl/openl3 (606 stars) with PyPI package and readthedocs docs; widely cited as a baseline. Strengths: dead-simple to use, permissive license, well-documented. Docked for: (1) benchmark claims are provider-reported and from 2019; (2) repo is in low-maintenance mode; (3) authors disclosed a temporal-overlap bug in the training code. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ICASSP 2019 paper (Cramer, Wu, Salamon, Bello; NYU MARL): self-supervised audio-visual correspondence embeddings
- · Repo marl/openl3 (606 stars), PyPI package, TF2 support; announcement via AUDITORY list May 2019
- · Benchmark claims are provider-reported and from 2019; repo in low-maintenance mode; authors disclosed a temporal-overlap bug in training code