Lucy
Talk
Voice OS · 2026

Wakeword Detection

A wakeword is the trigger phrase that takes a voice AI from passive listening to active conversation. Wakeword detection is what makes always-on voice AI possible without uploading every second of audio to the cloud. The detector runs continuously on a tiny neural net, typically a few hundred kilobytes, that is optimized for one task: deciding whether the most recent few seconds of audio contain the wake phrase. Everything else gets dropped immediately. Wakewords are the privacy and bandwidth foundation of always-on voice.

WHAT TO LOOK FOR

The three things that actually matter

1

On-device inference

The wakeword model runs in the browser or on the device's CPU, never the cloud. Audio that does not contain the wakeword is discarded immediately and never transmitted. This is the privacy guarantee that makes always-on listening acceptable to users.

2

False accept and false reject rates

A wakeword detector is judged on two error rates: how often it triggers on non-wakeword audio (false accept) and how often it misses an actual wakeword (false reject). Production wakewords aim for one false accept per 24 hours and one false reject per 100 attempts.

3

Custom wakeword training

Off-the-shelf wakewords like 'Hey Siri' or 'Alexa' are trained on millions of utterances. Custom wakewords for individual products are typically trained on tens of thousands of crowd-sourced recordings, which is why custom wakewords often have higher error rates than the brand defaults.

TLDR:Lucy OS1 supports an optional wakeword for hands-free activation. The detector runs locally in the browser using a small neural net, so no audio leaves the device until the wakeword fires. Once it does, the audio session opens and the standard pipeline takes over. For users who prefer click-to-talk activation, the wakeword can be disabled. For users who want truly hands-free operation while cooking, driving, or working, the wakeword turns Lucy into an always-available presence without the privacy cost of streaming audio continuously.

Why Lucy OS1

On-device inference

The wakeword model runs in the browser or on the device's CPU, never the cloud. Audio that does not contain the wakeword is discarded immediately and never transmitted. This is the privacy guarantee that makes always-on listening acceptable to users.

False accept and false reject rates

A wakeword detector is judged on two error rates: how often it triggers on non-wakeword audio (false accept) and how often it misses an actual wakeword (false reject). Production wakewords aim for one false accept per 24 hours and one false reject per 100 attempts.

Custom wakeword training

Off-the-shelf wakewords like 'Hey Siri' or 'Alexa' are trained on millions of utterances. Custom wakewords for individual products are typically trained on tens of thousands of crowd-sourced recordings, which is why custom wakewords often have higher error rates than the brand defaults.

Speaker verification

Optional second stage that verifies the wakeword was spoken by the device owner, not a guest or a TV. Reduces false accepts to near-zero in shared environments at the cost of a brief enrollment step.

Latency from wake to active

The time between the wakeword finishing and the audio session being ready to accept commands. Sub-300 millisecond wake latency feels seamless; longer than that and users either repeat the wakeword or start speaking before the session is ready.

Privacy LED or chime

Visible or audible confirmation that the wakeword fired and the audio session is open. This is both a usability cue and a privacy guarantee that the user knows when audio is being captured.

QUICK COMPARISON

Lucy OS1 vs most AI tools

Capability Lucy OS1 Most AI tools
Memory across sessions ✓ Permanent, never resets ✗ Resets after every session
Voice quality ✓ Lucy OS1 Natural Voice (best-in-class) ✗ Basic STT, struggles with noise
Calendar awareness ✓ Reads Google Calendar in real time ✗ No calendar access
Available 24/7 Always on, any device Available but stateless each time
Gets personal over time ✓ Builds your context continuously ✗ Starts from zero every session

Try Lucy OS1, setup takes 30 seconds

Voice-first AI with memory and calendar integration. Free to try.

Start Talking

Free tier available. No credit card required.

GET STARTED

How to use Lucy OS1

1

Create your free account

No credit card required. Sign in with your Google account and you're inside in under a minute.

2

Connect your Google Calendar

Lucy reads your upcoming events before every conversation, so it already knows your day before you say a word.

3

Start talking about wakeword detection

Speak naturally. Lucy listens, responds by voice, and begins building context from your very first exchange. The more you use it, the better it gets.

Start for free → Free tier available. No credit card.

Frequently Asked Questions

Does always-on listening mean my voice is uploaded continuously?
No, when implemented correctly. The wakeword detector runs locally and discards all audio that does not match. Only after the wakeword fires does any audio leave the device. Vendors that follow this pattern are not recording your conversations.
Can wakeword detection happen in a browser?
Yes. WebAssembly runtimes for small neural nets make wakeword detection feasible in the browser at low CPU cost. Lucy OS1 uses this approach to support hands-free activation in Chrome, Safari, and Firefox without a native app.
Why do wakewords sometimes trigger on TV or radio?
Off-the-shelf wakewords are trained to recognize the phrase regardless of speaker, so a voice on TV saying 'Hey Siri' will trigger a real Siri device. Speaker verification mitigates this by requiring the trigger to match the device owner's voiceprint.
What is the latency cost of wakeword detection?
The detector itself adds 50 to 100 milliseconds of latency between the wakeword being spoken and the session opening. This is below human perception, so the system feels instant. Custom wakewords with larger models can add more.
Are wakewords required, or is push-to-talk fine?
Push-to-talk works well for stationary use at a desk. Wakewords are required for truly hands-free use cases like driving, cooking, or workouts where pressing a button is impractical. Most production voice OSes support both.
Can wakeword models be customized to my voice?
Some implementations support speaker enrollment, which adapts the detector to your voice in 10 to 30 seconds of recording. This both reduces false accepts on other voices and improves recognition of your own voice in difficult acoustic conditions.

MORE IN THIS CATEGORY

→ Voice OS Architecture → The Voice AI Audio Pipeline → Voice AI Latency Budget → Endpointing in Voice AI → Barge-In and Interruption Handling → The Memory Layer of a Voice OS → Voice OS Context Window → Voice OS Permissions Model → See all

COMPARE LUCY OS1

Lucy OS1 vs Siri → Lucy OS1 vs ChatGPT → Lucy OS1 vs Google Gemini → Lucy OS1 vs Google Assistant → Lucy OS1 vs Amazon Alexa → See all comparisons →

Welcome