Build on-device AI call assist that answers calls, identifies who's calling and why using a local LLM, and gives users smart action options — entirely on-device, no call audio leaves the phone. Licensable SDKs for Android, iOS, Linux, and embedded.
On-device AI call assist combines Cheetah Streaming Speech-to-Text, Orca Streaming Text-to-Speech, picoLLM Inference, and Rhino Speech-to-Intent in a single local pipeline. Orca greets callers, Cheetah transcribes what they say in real time, picoLLM analyzes what's said and declines the call if it's suspicious. If it's not suspicious, Rhino captures the phone owner's decision — i.e., intent to direct Orca to respond accordingly. Most implementations require the user to select a source language upfront or still route at least one stage through a cloud API. This pipeline eliminates both constraints: every stage runs on the device.
Cheetah Streaming Speech-to-Text beats Google Cloud STT in word error rate and word emission latency across all tested languages, and outperforms Azure STT in several benchmarks, per the open-source real-time transcription benchmark — even before it's customized for the use case. It emits words at 590ms median latency, typically one word behind the speaker. Cheetah requires less compute than any other local engine tested.
picoLLM Compression quantizes language models so they run on phones, browsers, and embedded boards with no cloud while preserving task accuracy. Its minimalistic memory usage allows multiple models to run alongside Cheetah, Rhino, and Orca during a live call without introducing any network latency or privacy issues.
Most voice command systems run a two-step pipeline: speech-to-text converts audio to a transcript, then a separate NLU model parses that transcript for intent. Every step accumulates error and compounds latency. Rhino Speech-to-Intent is an end-to-end engine with a single model that maps spoken audio directly to a structured intent with typed slot values. Higher accuracy even in noisy environments. No hallucinations. No intermediate transcript.
Most high-quality TTS solutions require hundreds of megabytes of RAM. Orca TTS uses 29 MB peak memory, 10–50× less than any other on-device alternative, except for ESpeak. This makes Orca the only natural-sounding TTS deployable in any environment, including browser tabs, mobile apps with strict out-of-memory limits, and embedded devices. Orca Streaming TTS is built for real-time LLM applications — it starts reading LLM responses as soon as the LLM creates a meaningful word or phrase.
Google Pixel Call Assist and Apple Call Screening are exclusive to Pixel and iPhone hardware. Picovoice SDKs let any OEM ship a comparable or more advanced AI call assist as a built-in dialer feature on Android or as a third-party iOS app. No Google or Apple dependency. No backend to operate.
STIR/SHAKEN authenticates the caller ID. AI call assist understands what the caller actually says and reasons about it. Carriers can deploy on-device call assist as a differentiated feature in their dialer app — better than any cloud-based answering bot, because it works even when the network is congested.
VoIP and business phone apps embed an on-device AI receptionist that screens unknown calls, identifies prospects, and presents action options. The same SDKs work on iOS even though Apple does not expose Call Screening to third parties. No cloud API costs or privacy issues.
Cloud AI receptionists like Goodcall, Smith.ai, and Rosie process customer audio through their own servers, requiring agreements under HIPAA, GDPR, and relevant regulations. Picovoice has no data controller relationship with your callers — caller audio stays on the device. Compliance reviewers approve faster.
A complete working recipe in Python. Open-source on GitHub. Runs 100% on-device.