Porcupine Wake Word

Porcupine Wake Word is a developer tool. Picovoice's on-device wake word engine for adding custom voice triggers to apps and hardware. Train keywords in seconds with no training data, then deploy the same model across Android, iOS, web, desktop, Linux and embedded targets with audio processed locally.

Porcupine Wake Word is an on-device keyword spotting engine from Picovoice that gives applications an always-listening voice trigger without sending microphone audio to a server. Developers define a keyword or short phrase, train a model in the self-service console, and ship it to embedded hardware, phones, browsers, desktops or servers. The runtime stays under 1MB and keyword files sit under 25KB, so multiple triggers can coexist in one product with little compute overhead. The training flow is the part that separates it from older open-source options. There is no dataset to collect and no ML pipeline to maintain: a type-to-train interface converts text into a production model in seconds, and the download is already optimized for the target platform. Sensitivity runs on a 0 to 1 scale so teams can trade missed detections against false alarms for a given room, device or microphone. English, German, Spanish, French, Italian, Japanese, Korean, Portuguese and Chinese keywords can run side by side in the same app at no extra runtime cost. In practice the engine is used by product teams adding voice activation to consumer devices, by industrial and field software that needs hands-free control, and by robotics or automotive groups where connectivity is unreliable. Picovoice publishes its own benchmark against Snowboy and PocketSphinx, reporting 97.3% accuracy at one false alarm per 10 hours and 0.6% CPU utilization on a Raspberry Pi 5 32-bit build, with a separately stated 97.1% figure under 10 dB SNR noise. Alternatives split into two camps: cloud assistants such as Alexa or Google, which are easy but require connectivity and raise privacy questions, and open-source engines, which are free but demand training data and tuning. Porcupine sits between them, and the trade-off is a commercial license plus a Picovoice account dependency. Buyers who need guaranteed performance on unusual hardware or accents can commission dedicated model engineering through an NRE engagement.