mwirth7/soundscape-pretrain

Soundscape Pretrain Dataset

Audio clips from the Australian Acoustic Observatory (A2O) and Arbimon / RFCx platforms, pre-processed for self-supervised pre-training of soundscape models.

Splits

Events are filtered based on spectrogram properties:

FLATNESS_MAX     < 0.564       # drop broadband noise (higher = less selective)
LOW_FREQ_MAX     < 0.868       # drop wind/rumble
TEMPORAL_CV_MIN  > 0.307       # drop steady noise (lower = less selective)
HIGH_FREQ_CV_MIN > 0.373       # drop constant tonal drones
SPECTRAL_SHARPNESS_MIN > 8.3   # drop dull spectra (lower = less selective)

These values have been deterimined to achive ~80% percision on all BirdSet soundscape test datasets.

Split	Source	Approx. clips
a2o	Australian Acoustic Observatory	~336 k
arbimon	Arbimon / RFCx	~749 k

Usage

from datasets import load_dataset

# Streaming (recommended for training)
ds = load_dataset("earthspecies/soundscape-pretrain", split="a2o", streaming=True)
for item in ds:
    audio = item["audio"]  # {"array": np.ndarray float32 mono, "sampling_rate": 32000}
    energy = item["segment_mean_energy"]
    break

# Full download
ds = load_dataset("earthspecies/soundscape-pretrain", split="a2o")

Fields

Field	Type	Description
audio	Audio(32 kHz)	Mono FLAC, 2–8 s at 32 kHz
source	string	"a2o" or "arbimon"
start_seconds / end_seconds	float64	Clip offsets (s) within the recording
segment_mean_energy	float32	Mean log energy of the segment
spectral_flatness	float32	Wiener entropy / spectral flatness
spectral_sharpness	float32	Spectral sharpness (high-frequency tilt)
low_freq_ratio	float32	Fraction of energy in low frequencies
temporal_cv	float32	Coefficient of variation of temporal energy
high_freq_cv	float32	Coefficient of variation of high-freq energy
recording_id	int64	A2O only: A2O recording ID
site	int32	A2O only: A2O site ID
project_id	string	Arbimon only: Arbimon project ID
stream_id	string	Arbimon only: Arbimon stream ID