Picovoice Launches Koala, a High-Quality Noise Suppression Engine with Raspberry Pi Support
New machine learning noise suppression engine aims to beat Mozilla's RNNoise without needing a hefty machine for local processing.
Picovoice has announced a new string to its audio-processing bow: Koala, a noise suppression engine that it claims offers high-quality on-device results even when running on embedded hardware like the Raspberry Pi range of single-board computers.
"A good noise suppression engine should," Picovoice claims, run on the edge for minimal latency. Latency over 200ms affects the sense of presence, and humans start talking over each other. [It should] increase intelligibility. Like any product, it should create value, i.e., enhance speech. [It should] be platform-agnostic. It should offer unified experiences across platforms, whether on mobile, web, desktop, or all. We could not find anything readily available that meets all three criteria above."
That's where Koala comes in. Designed for use with the company's voice recognition engines, though also usable on its own, Koala is designed to process all audio data on-device with higher quality than the open-source RNNoise from Mozilla — with Picovoice claiming a four- to fivefold improvement in removing unwanted background noise from audio streams, based on an in-house benchmark it has developed and released for all to try.
The launch comes around a year after Picovoice released its second-generation on-device speech-to-text engines, Cheetah and Leopard — the former offering real-time transcription while the latter runs only on pre-recorded snippets in exchange for a lower word error rate (WER) of 11 per cent to Cheetah's 14.34 per cent.
Picovoice is making Koala available using Python, C, and JavaScript application programming interfaces (APIs), along with Android and iOS support. These run on Linux and Windows on AMD64/x86_64 hardware, on macOS under AMD64/x86_64 or Arm64 hardware, and on the Raspberry Pi 3, Raspberry Pi 4, and NVIDIA Jetson Nano embedded devices.
As with its other engines, while the software runs on-device it requires internet connectivity to validate a license key — with the company's free plan giving users 25 hours a month, down from the company's earlier 100-hours-a-month offering, across Leopard, Cheetah, Koala, and its Octopus speech-to-index engines.
More information on Koala is available on the company's website, including an in-browser demo; demo source code has been published to GitHub under the permissive Apache 2.0 license.
