Polyn Technology Ltd. (Tel Aviv, Israel), developer of the Neuromorphic Analog Signal Processor (NASP), has announced the availability of NeuroVoice, which performs on-chip voice extraction from noisy backgrounds.
The NASP chip design is optimised for a sparse neural network, with only the necessary connections between neurons required for inference. Determining those sparse networks is done using Polyn development tools. The company claims this brings a power consumption advantage over some compute-in-memory architectures, which connect each node to each neighboring node regardless of the neural network being accelerated.
Polyn claims to be the first company to implement the combination of voice detection (VD), and voice extraction (VE) within analog neuromorphic silicon.
NeuroVoice operates within a 100 microwatt power consumption budget and performs inference in 20 microseconds. As a result it can offer clear voice and immediate adaption to any background conditions.
“Earbuds, smartphones, hearing assistance, gamer headphones, and intercom systems need new technology to bring voice processing to a new level,” said Eugene Zetserov, vice president of marketing and business development at Polyn, in a statement.
“Current methods of voice signal processing are power hungry and, in some cases, fall short. Immediate voice recognition is important for hearing assistance devices. AI-based extraction of the voice signal in a noisy environment, including irregular noises, provides a better hearing experience than standard noise cancellation filters. A neural network is the perfect tool for voice processing and Polyn offers it on a tiny neuromorphic analog chip.”
Polyn framework and development tools make it possible to customize the existing chip to provide additional voice features such as wake word detection (WWD) and keyword spotting (KWS).
Related links and articles: