Combine on-device wake word, speech-to-intent, and streaming text-to-speech so users call any contact by voice, including international names and freshly added contacts. Custom voice models are generated per user via cloud API and run entirely on-device.
The on-device voice dialing pipeline runs three Picovoice SDKs in a loop on the user's device: Porcupine Wake Word, Rhino Speech-to-Intent, and Orca Streaming Text-to-Speech. Porcupine listens for the wake phrase. Rhino captures the callContact intent with contact, company, and phone-type slots directly from natural speech. The app looks up the contact against a local address book, dials the selected number, and Orca speaks the confirmation. If the name is ambiguous or the requested phone type is missing, the pipeline loops with a spoken follow-up. Audio and contact data never leave the device.
Porcupine Wake Word activates the app when the user utters the chosen phrase, such as "Hey Pico". Product teams can train a branded wake word in seconds in the Picovoice Console or via the Porcupine Model Training API and deploy it across Android, iOS, embedded Linux, and Raspberry Pi. Porcupine runs always-on at very low CPU and battery cost, so the rest of the pipeline only spins up when it needs to, which is essential for battery-powered devices, such as mobile phones, a pair of smart glasses, or a Bluetooth headset.
Rhino Speech-to-Intent captures the callContact, selectContact, and selectPhone intents with typed slots: the contact name, the company (to disambiguate "Sarah at Acme" from "Sarah at Cloudway"), and the phone type (mobile, work, home). The Rhino context is built dynamically from the user's local contact list at initialization, so the model recognizes exactly the names in the address book. Most voice command systems chain speech-to-text with a separate NLU model; each step accumulates error and compounds latency. Rhino infers intent and typed slot values directly from audio, holding higher accuracy in noisy environments, which is critical in a moving car or a busy sidewalk.
Orca Streaming Text-to-Speech reads each dialer response aloud, such as "Calling Sarah Chen on mobile," "I found Sarah Chen and Sarah Khan. Which one?", or "There is no work number for Sarah Chen. Should I try mobile?". This way the user never has to look at the screen while driving, walking, or operating equipment. Most high-quality TTS engines require hundreds of megabytes of RAM. Orca uses 29 MB peak memory, 10–50× less than any natural-sounding on-device alternative, which fits easily inside a car head unit, a Bluetooth headset firmware image, or a pair of smart glasses. First-token latency is 128 ms, fast enough that spoken confirmations feel conversational.
Automotive OEMs and tier-one suppliers can embed voice dialing directly in the head unit, ADAS voice UI, or steering-wheel button flow. No dependency on Apple CarPlay Siri, Android Auto, or a cloud voice partner. Hands-free driving laws in the US, EU, and Australia treat voice dialing as compliant driver interaction; on-device execution keeps the pipeline responsive even when cellular drops in tunnels, canyons, or dead zones.
Smart glasses, smart watches, and hearables have single-digit-megabyte memory budgets and no room for cloud round-trips. The three-SDK pipeline (Porcupine + Rhino + Orca) weighs a few megabytes total and runs on embedded Linux and RTOS-class hardware. The wake word triggers on the wearable, the wearer says "call Alex," and the paired phone places the call using its local contact list. No voice data is shared with a cloud or a voice partner.
For users who cannot easily use a touchscreen (visually impaired, motor-impaired, elderly users, or users in rehabilitation), voice dialing is the primary way to place a call. The on-device pipeline works without an internet connection, without a cloud account, and without a data plan, which matters for accessibility deployments in care facilities and low-bandwidth regions.
Enterprise fleet safety programs enforcing hands-free driving compliance can embed voice dialing into the fleet app on rugged Android handhelds. VoIP and business phone apps (OpenPhone, Dialpad, Grasshopper, Aircall, RingCentral) can add voice-first dialing on iOS and Android. The same SDKs work on iOS even though Apple does not expose Siri contact calling to third parties.
A complete working recipe in Python. Open-source on GitHub. Runs 100% on-device.
callContact intent with contact, company, and phone-type slots. If the contact name matches multiple entries (for example, Sarah Chen at Dyson and Sarah Khan), the app asks a spoken follow-up ("I found Sarah Chen, Sarah Khan. Which one?") and captures the answer as a selectContact intent. The same pattern handles phone-type disambiguation (mobile, work, home) via the selectPhone intent. Disambiguation also works when contact names come from different languages or scripts, because the Rhino context is built from the caller's own list. See the end-to-end intent inference approach for why this is more accurate than STT-plus-NLU stacks.