By the Outspoken Team · October 22, 2026
Wake Word Training Data: How Much You Need and Where It Comes From
Every wake word model is only as good as the audio it was trained on. Before you can train anything, you need a dataset of your wake word spoken many different ways, plus a much larger set of "not the wake word" audio to teach the model what to ignore. This is wake word training data, and getting it wrong is the most common reason a model misses activations or fires on nothing.
This guide covers where training data comes from, how much you actually need, and why synthetic data has mostly replaced manual recording for custom wake words.
The Two Datasets Every Model Needs
Positive examples — recordings of someone actually saying your wake word. The model learns the acoustic pattern from these.
Negative examples — everything else: speech, music, silence, background noise, other wake words. Without a large and varied negative set, a model can't tell your wake word apart from normal conversation, and you get false activations.
A dataset that's all positives and no negatives isn't a wake word dataset — it's a party trick. The negative set is usually 10-100x larger than the positive set, because "everything that isn't the wake word" is a much bigger space than "the wake word."
Where Training Data Comes From
Manual Recording
The original approach: record yourself (and ideally other people) saying the wake word dozens or hundreds of times, in different rooms, at different distances, with different moods. This is how early tools like Snowboy worked.
The problem: it doesn't scale. A few hundred recordings from one or two voices gives you a model that recognizes those specific voices in that specific room. It won't generalize to a different speaker, accent, or microphone.
Synthetic Speech (TTS)
Modern wake word training — openWakeWord, microWakeWord, and Outspoken all use this approach — generates positive examples with a text-to-speech engine instead of a microphone. A model like Piper TTS synthesizes your wake word thousands of times using many different voice models, speaking rates, and pitch variations.
This solves the scale problem: instead of recording 300 clips yourself, you generate 10,000+ synthetic clips covering far more vocal variation than any one person (or even a small recording session with volunteers) could produce in a reasonable timeframe.
Negative examples come from large public speech and audio datasets — general conversation, podcasts, ambient noise, music — mixed in during the next step.
Augmentation
Raw synthetic clips sound too clean. Real-world microphones pick up room reverb, background noise, and distance from the speaker. Augmentation simulates this by layering the synthetic speech with:
- Background noise (kitchen, traffic, TV, HVAC hum)
- Room impulse responses (simulated reverb for different room sizes)
- Volume and distance variation
A model trained only on clean synthetic speech will perform well in a silent testing room and fall apart in an actual kitchen or living room. Augmentation is what makes the dataset reflect how the model will actually be used.
This is what Outspoken automates
When you train on Outspoken, you don't record or upload anything. Entering your wake word triggers the full pipeline: TTS generation, noise and reverb augmentation, and training against a shared negative dataset — the same approach openWakeWord uses, run on managed GPU infrastructure.
How Much Training Data Do You Actually Need
There's no universal number — it depends on the wake word itself — but ranges from real training runs look like this:
| Positive Examples | Result |
|---|---|
| 100 – 1,000 | Fast to train, but misses unusual pronunciations and accents |
| 2,000 – 10,000 | Good balance for most wake words — this is where most production models land |
| 10,000 – 25,000 | Maximum variation; worth it for short or phonetically ambiguous wake words |
Two things matter more than raw volume:
- Voice diversity, not just clip count. 10,000 clips from 5 TTS voices generalize worse than 10,000 clips from 50 voices. This is why synthetic generation with many TTS models outperforms manually recording the same person 10,000 times.
- The wake word's phonetics. A short, soft wake word ("Go", "Hey") needs disproportionately more data than a longer, acoustically distinctive one ("Hey Jarvis") to reach the same false-activation rate — and sometimes no amount of data fixes a weak phrase. See the training parameters guide for the phonetics side of this.
Do You Need to Collect Your Own Data?
For almost all custom wake word use cases: no. Synthetic TTS generation plus augmentation, done at sufficient scale, outperforms what most teams could realistically collect by recording volunteers. Manual recording only makes sense if you need a wake word to work for one specific, known speaker and nobody else — a narrow use case most products don't actually have.
If you're evaluating tools, compare how each one handles data generation — it's one of the biggest practical differences between them, not just pricing or platform support.
Ready to see this pipeline run on your own wake word? Sign up for Outspoken — training data generation is included, first model is free.