Picovoice's free tier is gone. Re-cost the always-on layer at 53 ms per second.

Picovoice's Porcupine, the on-device wake-word SDK a lot of hobby and prototype voice stacks were built on, had its free tier deprecated as of 30 June 2026, replaced by a 7-day trial followed by paid, according to a post published on dev.to by the vendor behind a competing detector ^[1]. That turns the cheapest layer of the stack into a line item, and since wake-word is the component that listens to the mic 24 hours a day and therefore dominates the energy and compute budget of a voice-first device ^[3], it is the layer worth re-costing first.

The budget has three tiers. Voice activity detection is the floor at roughly 1% CPU and only answers whether speech is present at all ^[4]. Wake-word answers whether a specific phrase was said, typically from a 10 to 100 KB model, and runs 1 to 10% CPU always-on ^[5]. Full ASR is a different order of magnitude: around a 60 MB model and 30 to 40% CPU on a mid-range Android, and it only fires after the wake-word hits ^[6].

The number now being offered as the bar is 5.3% of one Cortex-A53 core on a $15 Raspberry Pi Zero 2 W, described as 24/7 sustained ^[7]. Expressed as a real-time factor, that is 53 milliseconds of CPU time per second of audio ^[8]. The board is a quad-core A53 at 1 GHz with 512 MB of RAM and no GPU ^[9], so the same load is about 1.3% of the machine's total four-core capacity ^[12]. The vendor's framing is that this leaves 95% of the core plus three other cores for the application, and that on a Pi 5 the detector is effectively free at under 1% ^[11].

Read the measurement conditions before you plan around it. The Pi Zero 2 W figure comes from a 60-second sustained live-mic push; the Pi 3, 4 and 5 numbers in the same table are estimates scaled by clock speed and architecture generation, not independently benchmarked ^[10]. A one-minute window is not the same evidence as a 24/7 claim, and it is the vendor's own measurement of the vendor's own model ^[24].

Accuracy is reported on 5,240 positive and 6,416 negative utterances ^[13], which is 11,656 in total ^[14]: ROC AUC 0.9966, PR AUC 0.9899, and at a 0.90 threshold precision 0.993 with recall 0.982 ^[13]. That precision implies roughly one false accept per 142 accepted detections ^[15]; the post rounds it to about one false positive per 100-plus triggers and about 2% of real wake-words missed ^[16]. The model is about 48 thousand parameters in a roughly 100 KB encrypted .vxrt file at 16 kHz mono with 32 ms frames ^[17], against Whisper base.en's roughly 74 million parameters ^[19]. Training data is 100% synthetic with augmentation and no human voice samples, which the vendor presents as removing a licensing question for commercial deployment ^[18].

The catch sits in the same place Picovoice's did. The free tier is one fixed phrase, "Hey Assistant"; custom phrases are paid, with the training pipeline available on request ^[20]. If what you actually needed was a custom wake word with no per-activation billing ^[25], this is a different paywall at the same tier. The open alternative, openWakeWord, is Apache-2.0 with community models, but the post concedes that training a custom phrase means gathering data, training and iterating ^[21], and building from scratch on TF Lite needs ML expertise and hours ^[22].

Watch for an independent reproduction of the 5.3% figure over hours rather than a minute, with thermals attached. Watch how custom-phrase pricing is structured, per device or per activation, since that is the variable that made the old tier hurt. And note that the model ships encrypted ^[17]; whether it needs a periodic activation check is the detail that decides whether "on-device" survives a dead uplink, and the post does not say.