VGGish
The classic CNN audio embedding: VGG-style network trained on YouTube-8M/AudioSet.
About
VGGish is one of the founding artifacts of modern audio deep learning: a VGG-inspired convolutional network trained on a large YouTube corpus (the work behind it is Hershey et al., ICASSP 2017, "CNN Architectures for Large-Scale Audio Classification"), producing compact 128-dimensional embeddings from log-mel spectrogram patches. For years it was the default "just give me an audio feature vector" model in research code.
Google released the model through the TensorFlow models repository, and the weights remain easy to find (TF checkpoints, PyTorch ports like torch-vggish-yamnet). It still works fine as a lightweight baseline, but go in with open eyes: it is a 2017 architecture, Google has not meaningfully updated it, and modern embeddings (PANNs, BEATs, MuQ) beat it on most benchmarks. Use it when you need something tiny, deterministic, and universally reproducible, not when you need the best numbers.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No current published benchmarks; a 2017-era baseline that modern models consistently outperform.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (verified, metrics-v1). Evidence: ICASSP 2017 paper (Hershey et al., Google); official model in the tensorflow/models AudioSet section (Apache 2.0); enduring presence in research code and PyTorch ports. Strengths: tiny, reproducible, permissive license, universally understood baseline. Docked for: (1) effectively unmaintained by Google; (2) outclassed by every modern embedding model; (3) no meaningful published benchmark of its own. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 3.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · Google (Hershey et al., ICASSP 2017 'CNN Architectures for Large-Scale Audio Classification'); 128-dim embeddings from log-mel patches
- · Official model in tensorflow/models AudioSet section; community PyTorch ports widely available
- · Effectively unmaintained by Google; outclassed by modern embeddings; a legacy baseline, not a current SOTA claim