Wake Word and Keyword Spotting Training Data: What You Actually Need
21
Wake Word and Keyword Spotting Training Data: What You Actually Need
Wake word and keyword spotting training data has a specific structure that separates it from general ASR training: many short positive utterances of the target phrase from diverse speakers and environments, plus a substantial hard-negative set of near-miss utterances that must not trigger. Get either side wrong and the product either false-rejects the user or false-accepts random speech. This post walks through the collection structure, the volume and diversity numbers, and the collection frontier for new wake phrases across languages and accents in 2026.
TL;DR
Wake word training needs two matched sets: positive utterances of the target phrase and a hard-negative set typically 3 to 10 times larger.
Consumer production tier typically requires 100,000 to 500,000 positive utterances per wake phrase; global cross-accent coverage pushes this into the millions.
Diversity across speaker demographic, environment, and device class matters more than raw utterance count.
Each new language or accent variant is effectively a new wake-phrase collection, not a translation.
Silencio's contributor network supports scoped wake-word programs with target-phrase capture and matched hard-negative sets in one delivery.
What is wake word and keyword spotting training data?
Wake word data trains a small always-on model that listens for a specific trigger phrase ("Hey Assistant") and activates a larger downstream system when it detects that phrase. Keyword spotting extends the same principle to short command vocabularies ("Turn on", "Play music", "Answer call"). The training data structure is fundamentally different from general ASR.
Positive set. Recordings of the target phrase or keyword, from many speakers, in many environments, on many devices. This is what the model learns to recognize.
Hard-negative set. Recordings of near-miss phrases and everyday speech that must not trigger. "Hey Assistant" needs to not trigger on "Hey Alexis", "Hey Amanda", or "Say assistant." Hard negatives are the specific failure surface.
Diversity axes. Speaker demographic (age, gender, accent), environment (in-cabin, in-home with TV, in-office, outdoor), and device (far-field smart speaker, near-field phone, in-car array microphone).
What volume of positive utterances do you need?
Depends on the false-reject and false-accept targets. Rough bands.
Product tier | Positive utterances per wake phrase | Notes |
|---|---|---|
Prototype (demo-quality) | 5,000 to 10,000 | Wake works in most conditions, not tuned for production |
Beta (early production) | 20,000 to 50,000 | Acceptable in target environments, still tunable |
Consumer production | 100,000 to 500,000 | Wide device and environment coverage, low false-reject in target markets |
Global consumer production (across accents) | 500,000 to 2,000,000+ | Cross-accent, cross-region coverage |
These are positive-utterance counts across the diversity axes above. Naive collection (10,000 utterances from 20 speakers in one environment) produces a model that works for those 20 speakers in that environment. Diverse collection (10,000 utterances across 500 speakers in 5 environments on 3 device categories) produces a model that generalizes.
What about the hard-negative set?
The hard-negative set is typically 3 to 10 times the size of the positive set, depending on how sensitive the wake word phrase is to confusion. Categories of hard negatives:
Phonetically similar phrases intentionally collected ("Hey Alexis" for "Hey Alexa")
Common everyday speech captured in target deployment environments
Background TV/radio dialogue as ambient acoustic negatives
Music with vocals as spectral negatives
The false-accept rate the product can tolerate drives how much hard-negative data is needed. A voice assistant that opens a microphone on false-accept has different tolerance from one that initiates a purchase transaction.
What are the per-accent and per-language collection challenges?
For a global wake phrase, the positive set has to cover every accent and language the product ships in. Concrete pattern:
Same wake phrase ("Hey Assistant" in English) across US, UK, Indian, Nigerian, Australian, Singaporean, and other accent regions
Translated wake phrase across every supported language (each language is essentially a new wake-phrase collection)
Sub-accent variation within languages (US Southern, US Northeast, US Midwest for US English; regional Nigerian English varieties, etc.)
Each accent-language combination typically needs 20,000 to 100,000 positive utterances to reach production tier, depending on how much the accent diverges from the base training accent.
What environments matter most?
The environments matter as much as the utterance count. For general consumer voice assistants:
In-home: kitchen with running water, living room with TV, bedroom quiet
In-vehicle: idle, urban driving, highway
In-office: conference room, open plan, private office
Outdoor: street, park, transit
For specialist deployments:
Automotive: full range of driving conditions, cabin acoustic profiles per vehicle model
Smart-speaker: full range of home rooms with typical background noise
In-ear: on-body acoustic conditions, walking, moving
Missing environment coverage in the training data shows up as environment-specific false-reject inflation in production.
What about device diversity?
Modern wake word detection has to work on:
Far-field: smart speakers (Amazon Echo, Google Home, HomePod, Baidu, Xiaomi, etc.)
Near-field: phones held in hand, phones on desk
In-cabin: automotive microphones with beamforming
On-body: earbuds and hearables with mic arrays
Each device class has different pickup characteristics. Training data captured on one class does not fully transfer to another.
Common misconceptions about wake-word training data
Three assumptions surface repeatedly in buyer conversations and are wrong.
Misconception 1: "Synthetic augmentation of a small positive set can substitute for real diversity." Pitch-shift, tempo-shift, and additive noise augmentation help at the margin but do not simulate real speaker variation across accent, age, and vocal quality. A model trained on augmented data from 50 speakers underperforms one trained on organic data from 5,000 speakers, at every accent that was not in the original 50.
Misconception 2: "Hard negatives can be generated from ASR corpora." Only for a subset. General ASR corpora contain some phonetically-adjacent utterances by coincidence, but the specific near-miss confusions for a chosen wake phrase have to be intentionally collected. "Hey Alexis" recordings do not appear in a general ASR corpus in useful volume; they have to be commissioned.
Misconception 3: "One wake phrase collection covers all languages via translation." Wake phrases translate at the semantic level, not the acoustic level. The English phrase and its Vietnamese, Arabic, or Swahili translation share nothing acoustically. Each new language is a full positive-and-negative collection program, not a translation task.
How does Silencio approach wake-word collection?
Silencio's contributor network can capture target-phrase utterances at scale across the demographic, environment, and device diversity required. For scoped wake-word programs, we can commission:
Specific target phrase collection across contributor demographics
Environment-tagged capture in the environments the wake phrase will deploy to
Multi-device capture where the same contributor records on different devices
Hard-negative collection tuned to the specific confusion set of the chosen phrase
Delivery in weeks depending on volume and language coverage.
Silencio's contributor network supports scoped wake-word collection across languages, accents, environments, and devices, with target-phrase capture and matched hard-negative sets in a single program.