Identifies who is speaking by voiceprint, personalizes responses, and grants or denies voice commands based on the speaker's role. Built for enterprises developing their own Alexa Voice ID or Google Voice Match.
An AI voice assistant with speaker recognition listens for a wake word, infers the user's intent directly from the follow-on spoken command, identifies who is speaking by voiceprint to grant or deny access based on the speaker's role, and responds to users all on-device with no cloud dependency. Porcupine Wake Word detects the wake word. Rhino Speech-to-Intent extracts the command as a structured intent. Eagle Speaker Recognition compares the speaker's voice against enrolled voiceprints and returns a similarity score. The assistant then checks the speaker's role (admin or user) against the command's permission level before acting. Orca Text-to-Speech speaks the response back.
Porcupine Wake Word provides always-on, low-power wake word detection for embedded devices. It listens continuously with minimal CPU usage and command pipeline only when the wake word is detected, keeping power consumption low between activations. Custom wake words can be trained in seconds using the Picovoice Console and exported for embedded, mobile, web, and desktop applications.
Rhino Speech-to-Intent extracts structured intents directly from spoken commands without an intermediate speech-to-text step, achieving a higher accuracy than Big Tech alternatives, like Google Dialogflow and Amazon Lex. For a speaker-aware voice assistant, Rhino's domain-specific models define which commands exist and what permission levels they require. You configure the intents (e.g., adminOnly, speakerPersonalized, generic) and slots in the Picovoice Console. Rhino only recognizes commands within the defined domain, eliminating out-of-domain misrecognition.
Eagle Speaker Recognition identifies who is speaking by comparing the speaker's voice against enrolled voiceprints. Eagle returns a similarity score for each enrolled speaker, and the application uses that score to determine identity and role. Enrollment takes a few seconds of speech. Eagle runs entirely on-device; voiceprint data is never transmitted to Picovoice or any third-party server.
Orca Text-to-Speech speaks the assistant's responses back to the user, such as "Hi Jane, playing your favourite playlist." "Admin command approved." or "Permission denied." after the access decision. Orca's 29 MB peak memory usage and 7MB model size make it well-suited for mobile and embedded applications.
Smart home manufacturers can restrict sensitive commands to enrolled household members. "Disarm the alarm" only works for verified adults, while children and guests can change the music and temperature without access to security or purchase functions.
Enterprises deploying shared kiosks, terminals, or workstations can assign different permission levels to different speakers. A floor manager can override production settings; an operator can only start and stop cycles. The device identifies the speaker by voiceprint in real time.
Only authorized clinicians can issue medication commands, adjust device settings, or access patient information by voice, giving clinicians hands-free authentication during procedures and consultations. All speaker recognition runs on the device, keeping voiceprint data off external servers.
Retailers can add voice commands to POS terminals and self-checkout kiosks where managers, store associates, and customers have different access levels. Voiceprints of employees are stored locally and never sent to Picovoice or any third party.
A complete working recipe in Python. Open-source on GitHub. Runs 100% on-device.
admin_similarity_threshold parameter lets you tune the tradeoff between security (higher threshold, fewer false accepts) and convenience (lower threshold, fewer false rejects).