Audio Analysis
Log-Mel Spectrogram
Turns each clip into a log-mel spectrogram — frequency bands over time, in dB — saved as a NumPy matrix (.npy) and as an image (.png) for image models such as CNNs.
When to use it
Feed spectrograms to your own image classifier, inspect recordings visually, or build a training set for a sound model.
In a pipeline
Audio File InputAudio Chunk / SlicerLog-Mel SpectrogramExport CSV
Settings
| Setting | Type | Required | Default or hint |
|---|---|---|---|
| Mel bands | number | — | 64 |
| Window (ms) | number | — | 25 |
| Hop (ms) | number | — | 10 |
| Lowest frequency (Hz) | number | — | 0 |
| Highest frequency (Hz, 0 = half the sample rate) | number | — | 0 |
| Dynamic range (dB) | number | — | 80 |
| Save the matrix (.npy, bands × frames) | checkbox | — | true |
| Save an image (.png) | checkbox | — | true |
| Image colours | select | — | magma |
Example
64 bands, 25 ms windows every 10 ms
- Mel bands=64
- Window (ms)=25
- Hop (ms)=10
| File Name | Mel Bands | Frames | Image Path |
|---|---|---|---|
| engine.wav | 64 | 501 | …/engine_logmel.png |
Worth knowing
- The .npy holds bands × frames as 32-bit floats: numpy.load(path) reads it back as is.
- In the image, low frequencies are at the bottom and loud is bright.
- Dynamic range floors everything more than that far below the loudest point, as librosa's top_db does.