Combine on-device wake word detection with speaker recognition so a phrase like "Alexa" wakes the device only when the enrolled user says it. Personalized, secure, and fully on-device.
“Hey Meta” to enroll a speaker“Hey Meta”Personalized wake words, combining always-on wake word detection with speaker verification, only trigger when the enrolled user says them — so identical phrases from different speakers are treated differently. Most speaker recognition engines require 2–3 seconds of continuous speech for reliable identification, with accuracy degrading significantly below one second. Wake words are far shorter, which is why only a few companies, such as Alexa Voice ID by Amazon and Personalized Hey Siri by Apple, can offer personalized wake words in production. This guide shows how Picovoice's Porcupine Wake Word and Eagle Speaker Recognition compose into a self-contained personalized wake word that can be deployed at scale.
Equal Error Rate (EER), where false accepts equal false rejects, measures speaker recognition performance. A lower EER is better. Across four wake words, Eagle leads. It achieves 0.00% EER on Jarvis and Snowboy (vs. Pyannote: 0.71% and 0.75%; SpeechBrain: 0.36% and 0.35%), scores 0.15% on Alexa (about 7× lower than both at 1.08%), and matches others at 0.38% on Smart Mirror. Overall, Eagle’s mean EER is 0.13%, about 5× lower than Pyannote (0.63%) and SpeechBrain (0.64%).
For a personalized wake word that needs to reject imposters reliably, Eagle offers production-grade accuracy at wake word length audio, allowing enterprises to build experiences similar to Google Voice Match, Apple's Personalized Hey Siri, and Amazon's VoiceID, with no dependency on Google, Apple, or Amazon.
| alexa | jarvis | smart mirror | snowboy | average | |
|---|---|---|---|---|---|
| EagleThis recipe | 0.15 | 0.00 | 0.38 | 0.00 | 0.13 |
| Pyannote | 1.08 | 0.71 | 0.38 | 0.36 | 0.63 |
| SpeechBrain | 1.08 | 0.75 | 0.38 | 0.35 | 0.64 |
Porcupine Wake Word is lightweight, accurate, customizable, and production-ready. With Porcupine, enterprises can train branded wake words and always-listening commands in seconds and deploy them across platforms. Industry-leading accuracy fits in even MCUs.
Eagle Speaker Recognition achieves a 0.18% equal error rate on VoxConverse, a widely used multi-speaker dataset containing real conversations across multiple languages, roughly 3× lower than SpeechBrain (0.49%) and 4× lower than Pyannote (0.70%). On wake word length audio, Eagle's accuracy improves further: a mean EER of 0.13% across four wake words, including two perfect 0.00% scores, as shown in the benchmark above. Eagle performs reliable speaker verification (1:1) and identification (1:N) on short audio with no audio sent to a cloud.
Smart speakers, displays, and home hubs serving multiple family members can route the wake word to the right user profile — kid asks for music, the hub plays from their playlist; parent asks for the calendar, the hub answers from theirs.
Earbuds and watches have single-digit memory budgets and zero room for cloud. Porcupine and Eagle fit in embedded devices. So only the wearer's wake word triggers, even when they share a kitchen with three other voices.
Cars need wake words that ignore the passengers. Combine Porcupine with Eagle speaker verification so only the driver's wake phrase triggers driving-relevant actions and brings driver settings, while passengers summon entertainment with a different or no profile.
Hands-free industrial HMIs, surgical ORs, medical bedside devices, and field tools need a wake word that won't fire on coworkers but only authorized personnel. Personalized and secure wake word gates the device to the operator on shift.
A complete working recipe in Python. Open-source on GitHub. Runs 100% on-device.