Koala is an on-device noise suppression engine that cleans speech audio frame by frame without sending it to a server. You create an engine instance, feed it consecutive PCM frames, and receive enhanced frames back in real time; a reset() call handles gaps between non-contiguous frames.
An AccessKey from the Picovoice Console authenticates each deployment. Validation contacts Picovoice license servers over the internet, while the audio processing itself runs entirely offline. Version 3.0.0 adds GPU and multi-core CPU execution and raises the minimum runtime versions to Node 18 and iOS 16.
Features
- Local processing: all voice enhancement runs on the device, with no audio leaving the machine
- Desktop platforms: Linux (x8664), macOS (x8664 and arm64), and Windows (x86_64 and arm64)
- Mobile and browser: Android, iOS, and Chrome, Safari, Firefox, and Edge via a web worker SDK
- Embedded: Raspberry Pi 3, 4, and 5
- SDKs: Python (pvkoala), Android (Maven), iOS (CocoaPods), C (pv_koala.h), and Web (@picovoice/koala-web)
- Frame API: a fixed frame length exposed by the engine, with process() returning an enhanced frame of the same size
- Demos: microphone and file demos for Python and C, plus Android, iOS, and web demo apps