Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

Hey Yossi Stop Word

Real recordings of the dismiss phrase "Go to sleep, Yossi", for training a microWakeWord stop model for the voice assistant "Yossi" (wake word "Hey Yossi", danielrosehill/Wake-Words). Linux Voice Assistant loads such a model with --stop-model and listens for it while it is speaking a reply, ringing a timer or making an announcement.

Recorded with the Voice Training Recorder (Project menu → Stop word) on a phone at 48 kHz / 24-bit mono FLAC, by more than one speaker, in ordinary rooms with ordinary noise. Kept apart from the wake word and ASR datasets: no take here is in any speech-to-text training set.

  • Positives (label positive): "Go to sleep, Yossi" (target go_to_sleep_yossi) in 14 deliveries (close, across the room, quiet, loud, fast, slow, tired, annoyed, firm, over speech or an alarm playing on another device, walking, after "Okay, thanks.") × 6 takes per speaker, and a separately labelled word-order variant "Yossi, go to sleep" (target yossi_go_to_sleep, 7 deliveries × 2).
  • Near-miss negatives (label negative, delivery near-miss, 3 takes each): "Go to sleep", "Time to sleep, Yossi", "Going to sleep", "Go to see Yossi", "Go to Yossi", "Yossi", "Hey Yossi" (the wake word, which must not fire the stop model), "Go to sleep, Yoni", "Go to sleep, Josie", "I'm going to sleep soon", "Go to the shop, Yossi", and a few ordinary sentences.

The dataset is empty until the first recordings are pushed, and it grows as more are recorded.

Columns

Column Meaning
audio (file_name) The take
text What was said
target The phrase the prompt belongs to: go_to_sleep_yossi or the variant yossi_go_to_sleep
label positive or negative (near-miss)
delivery How it was said: close, across, quiet, over-speech, over-alarm, after-reply, near-miss, ...
rep, instruction, prompt_id Repetition, the note shown to the speaker, the prompt
speaker, speaker_id, speaker_gender Who read it
device, recorded_at The recording device and time (UTC)
room, conditions, noise_flags Where, and what noise was present
peak_dbfs, floor_dbfs, snr_db, clipped_samples, ... Level and noise measured on the phone

Split: about 1 in 10 prompts held out as test, by prompt id hash.

Use

from datasets import load_dataset
ds = load_dataset("danielrosehill/Hey-Yossi-Stop-Word")
pos = ds["train"].filter(lambda r: r["target"] == "go_to_sleep_yossi" and r["label"] == "positive")

Downsample to 16 kHz for microWakeWord. These real clips are meant to sit beside synthetic (TTS-generated) positives and to be the held-out test that shows whether the model works on real voices.

Licence

CC BY 4.0, as the wake word dataset.

Downloads last month
23

Models trained or fine-tuned on danielrosehill/Hey-Yossi-Stop-Word