Audio Analysis
Audio Event Classifier
Names the sounds in each clip — glass breaking, dog barking, siren, engine knocking, speech, music — with the 527 AudioSet labels of the Audio Spectrogram Transformer, and a score for each.
When to use it
Security and safety monitoring, sorting a sound library, finding the clips with machinery faults, or flagging music in recordings that should be speech.
In a pipeline
URL / Stream IngestAudio Event ClassifierExport CSV
Settings
| Setting | Type | Required | Default or hint |
|---|---|---|---|
| Results | select | — | top |
| Labels to keep | number | — | 5 |
| Lowest score kept (0–1) | number | — | 0.1 |
| Window (seconds) | number | — | 10 |
| Model (Hugging Face id) | text | — | MIT/ast-finetuned-audioset-10-10-0.4593 |
| Run on | select | — | auto |
| Time limit (seconds) | number | — | 900 |
Example
The top labels of each clip
- Results=One row per clip: the top labels
- Labels to keep=3
| File Name | Label | Score | Labels |
|---|---|---|---|
| cam-04.wav | Glass | 0.91 | Glass (0.91); Shatter (0.84); Speech (0.22) |
Worth knowing
- One row per time window shows when each sound happens; a clip's own top labels take the highest score over its windows.
- A different Hugging Face audio-classification model (your own fine-tune, say) can be named under Model.
- Models run on the server in FlowDoo's Python and are downloaded once, on first use, into the data folder. A GPU is used when the server has one.
- A preview runs the model on the first 3 clips only; a run does them all.