AigizK/homai_wake_word_1020ms

Homai Wake Word 1020 ms

Binary training corpus for a wake-word detector that should react to `Һомай` and `Хомай`, and reject other speech. Every audio value is mono 16 kHz FLAC with exactly 16,320 samples (1020 ms).

Audited size

Source rows in the release: AigizK/Homai-Wake-Word (5,486), AigizK/homai_wake_word_omnivoice (830,772), and AigizK/bashkort_voice (928,519).

Construction

• —Positives at or below 1020 ms bypass ASR and receive leading silence.
• —Longer positives are localized with `AigizK/GigaAM-Bashkir-CV25-ONNX`. A row is kept only when its single recognized word interval plus acoustic margins fits in 1020 ms; otherwise it is rejected rather than repaired.
• —Every accepted positive contributes its unaugmented window plus three deterministic, distinct random augmentations drawn from gain (-20 to +9 dB), real background noise (5-25 dB SNR), time stretch (0.9-1.1, pitch preserved), pitch shift (±1-4 semitones, duration preserved), and speed perturbation (0.9-1.1, duration and pitch changed).
• —Negatives are not augmented. Short rows receive leading silence; long rows use the highest-energy 1020 ms window.
• —Split assignment is deterministic 90/5/5 by group_id; all variants and the two OmniVoice language renderings from one cloned source voice stay together.

Noise and licenses

Real-noise augmentation uses the noise portion of MUSAN (CC BY 4.0) and the real point-source noise recordings from RIRS_NOISES (Apache 2.0). The resulting dataset is published as license: other because downstream users must respect the licenses and attribution requirements of every source corpus.

The source corpora and their licenses remain authoritative. Provenance and exact augmentation parameters are included per row; the full build and rejection reports are stored in reports/.