Picovoice Cheetah

Cheetah is the go-to when audio cannot leave the device. Combined with Picovoice's wake-word engine Porcupine and intent-detection tool Rhino, it forms a complete offline voice stack that runs on hardware as constrained as a Raspberry Pi or an Arm-based microcontroller.

For developers building mobile apps that need voice input but can't send audio to the cloud (medical apps, in-car systems, certain government and defence work), there is no real alternative at this price point. The SDKs cover iOS, Android, Linux, Windows, macOS, Raspberry Pi, and microcontroller targets like the STM32 series.

That breadth of platform support is the second selling point: build once with Picovoice's API, deploy to phones and embedded devices alike. The trade-offs are honest. Accuracy on Cheetah is lower than Whisper-large or Deepgram Nova running in the cloud, because the model has to be small enough to fit on a phone or smaller device.

For dictation or transcript-grade work it's not the right pick. For voice commands, search inputs, and shorter utterances, it's plenty accurate. Commercial licensing is per-device on production tiers, which can add up at scale but is competitive against alternatives.

Language list is smaller than Whisper, mostly major European and Asian languages. The free tier for personal projects removes the friction from initial testing, and the SDK quality is consistently high across platforms.