project/
wav/clip01.wav
lbl/clip01.h5
manifests/train.tsv
manifests/valid.tsvWAV: the recorded signal
HDF5: sample-index event intervals
TSV: files in each split
train.tsv · first line: audio root · remaining fields separated by a literal TAB
/datasets/project
wav/clip01.wav 80000
wav/clip02.wav 80000
The row means 80,000 waveform samples—not 80,000 model frames. Set task.data to the manifest directory.
Toy vocabulary: task.unique_labels = ["call_A", "call_B"]
8 kHz audio; call_A from 1.00 s up to (not including) 1.15 s, call_B from 2.00 s up to (not including) 2.5 s; Both focal
start_frame_lbl = [8000, 16000] # a list holding the start indices of all events
end_frame_lbl = [9200, 12000] # a list holding the end indices of all events
lbl_cat = [0, 1] # a list holding the class indices of all events
foc = [1, 1] # a list indicating the focal status of all events (binary)
start_time_lbl = [1.0, 2.0] # a list holding the start times of all events in seconds
end_time_lbl = [1.15, 2.25] # a list holding the end times of all events in seconds
lbl = ["call_A", "call_B"] # a list holding the human-readable names for all eventsSame length. Starts inclusive; ends exclusive. Label indices follow the configured vocabulary.
A fully reviewed no-event file can have empty arrays. Missing annotations are known negatives. Unreviewed wav files have no HDF5 label files