augmentation
birdnet_stm32.audio.augmentation
¶
Data augmentation for audio and spectrograms.
Implements SpecAugment (frequency/time masking): apply_spec_augment on the loader's
spectrograms (librosa, hybrid), spec_augment_tf in the graph after the raw frontend,
which computes its spectrogram inside the model. Mixup works on whole batches in
TensorFlow ops: birdnet_stm32.data.generator.mixup_batch.
apply_spec_augment(spectrogram, freq_mask_max=8, time_mask_max=25, num_freq_masks=2, num_time_masks=2)
¶
Apply SpecAugment (frequency and time masking) to a spectrogram.
Zeroes out random contiguous bands along the frequency and time axes. Operates on a single spectrogram of shape [F, T] or [F, T, 1].
Reference: Park et al., "SpecAugment", 2019.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
spectrogram
|
ndarray
|
Input spectrogram [F, T] or [F, T, 1]. |
required |
freq_mask_max
|
int
|
Maximum width of each frequency mask (bins). |
8
|
time_mask_max
|
int
|
Maximum width of each time mask (frames). |
25
|
num_freq_masks
|
int
|
Number of frequency masks to apply. |
2
|
num_time_masks
|
int
|
Number of time masks to apply. |
2
|
Returns:
| Type | Description |
|---|---|
ndarray
|
Augmented spectrogram with same shape as input. |
Source code in birdnet_stm32/audio/augmentation.py
spec_augment_tf(spec, freq_mask_max=8, time_mask_max=25, num_freq_masks=2, num_time_masks=2)
¶
SpecAugment in TensorFlow ops on a batch [B, F, T, C], each sample masked independently.
The raw frontend computes its spectrogram inside the model, so the loader's
apply_spec_augment never sees it (every raw model up to 2026-11 trained without
SpecAugment). This runs on the frontend output during training only.
TensorFlow is imported here, not at module level: the loader workers import this
module for apply_spec_augment and must stay free of TensorFlow.
Masked cells take their band's median over time in the same sample (per channel), not
zero: the raw frontend maps silence to a band-dependent floor (median ~0.23 on
N6-B0), so zeros would be a pattern no real input produces. Mask widths are drawn
uniformly from [0, max), as in apply_spec_augment.