Data design
How do you collect wake word training data?
Updated 2026-08-01 · 4 min read
Short answer
A wake-word corpus needs three parts: positives — the phrase spoken by many speakers at many distances, angles, volumes and accents; negatives — hours of ambient household, street and media audio containing no wake word; and hard negatives — phonetically similar phrases that must not trigger. The false-accept rate is set almost entirely by the hard negatives, and the false-reject rate by speaker and distance diversity in the positives. Indian deployments need Indian-accented positives across languages, since a wake word spoken with a Tamil or Bengali phonology is acoustically different from the same word in the training set of a US-built model.
Key takeaways
- Hard negatives control false accepts; nothing else does.
- Positives need distance, angle and volume diversity more than raw count.
- Accent coverage matters even when the wake word is English.
Positives
Collect the phrase from a large speaker pool across ages and language backgrounds, at multiple distances and angles from the device, at whisper, normal and raised volume, and in the rooms where the device will live.
Negatives
Long-duration ambient recordings from real homes, shops and streets, including television and radio, which are the most common source of spurious triggers in Indian households.
Hard negatives
Systematically generate phonetically near phrases — same syllable count, similar onsets, rhyming variants — and collect them from the same speaker pool. This is the part most teams skip and the part that determines whether the device wakes up during dinner.
Evaluation
Report false accepts per hour of ambient audio and false rejects per hundred attempts, separately per accent group and per distance band. Aggregate numbers hide the accent that fails.
Frequently asked questions
How do you collect wake word training data?
A wake-word corpus needs three parts: positives — the phrase spoken by many speakers at many distances, angles, volumes and accents; negatives — hours of ambient household, street and media audio containing no wake word; and hard negatives — phonetically similar phrases that must not trigger. The false-accept rate is set almost entirely by the hard negatives, and the false-reject rate by speaker and distance diversity in the positives. Indian deployments need Indian-accented positives across languages, since a wake word spoken with a Tamil or Bengali phonology is acoustically different from the same word in the training set of a US-built model.
Positives?
Collect the phrase from a large speaker pool across ages and language backgrounds, at multiple distances and angles from the device, at whisper, normal and raised volume, and in the rooms where the device will live.
Negatives?
Long-duration ambient recordings from real homes, shops and streets, including television and radio, which are the most common source of spurious triggers in Indian households.
Hard negatives?
Systematically generate phonetically near phrases — same syllable count, similar onsets, rhyming variants — and collect them from the same speaker pool. This is the part most teams skip and the part that determines whether the device wakes up during dinner.
Related reading
Turn this into a dataset specification
Tell us the languages, speaker count and minutes. You get a written scope, a protocol and a fixed price within one working day.