Auto-labels the selected samples: every isolated word (or close two-word
phrase) becomes an extraction candidate labelled with its own
transcript. Review, edit labels, then upload — nothing is saved
without your check.
Words matching the blacklist tint red
and are skipped; everything else tints
green.
Listen before uploading — a confidently wrong transcript is the one
mistake no filter catches.
0%
Review & upload
Noise suppression over the selected samples.
model not loaded yet
Normalize peak:
to −dBFS
0%
Before / after
Splits each selected sample on silence, transcribes segments with the local model, and
tints segments containing the keyword
green /
others red.
Matches can then be extracted as new, keyword-labelled samples.
Keyword highlights
0%
dBFS
Peak-normalizes each selected sample to the target level and lets you keep the
results as new samples. (Also available inline as a pre/post option on the
Denoise tab.)
0%
Before / after
Grow the training set with the augmentations production speech pipelines use:
real background noise mixed at a controlled SNR, speed perturbation, room reverb,
and small time/gain jitter — all waveform-domain, complementing Edge Impulse's
training-time (spectrogram) augmentation. Keep EI's augmentation toggle
ON.
Loading labels…
WAV / MP3 / OGG — used only as mixing
material, never uploaded.
To avoid validation leakage, set the train/validation split in Edge Impulse to
group by the sourceSampleId metadata key (advanced training settings)
so augmented copies never validate against their own source.