Polyn NeuroVoice: Ultra-Low-Power Voice Processing Put to the Test

Real-time voice processing is becoming increasingly important in devices such as earbuds, hearing aids, smart microphones and other compact electronics. However, achieving useful audio processing while maintaining low latency and minimal power consumption remains a challenge, particularly in battery-powered applications.

Polyn Technology is addressing this challenge with NeuroVoice, its family of ultra-low-power neuromorphic analogue signal processors designed for on-device voice processing. Rather than relying on conventional, power-hungry digital processing architectures, NeuroVoice uses analogue neural networks to perform audio AI tasks close to the sensor.

A Range of Voice Processing Capabilities

The NeuroVoice family is designed to support several functions, including voice activity detection, speaker recognition, voice extraction, wake word detection and keyword spotting.

These capabilities enable devices to distinguish speech from background noise, identify individual speakers and respond to specific voice commands. Voice extraction is particularly relevant in noisy environments, where improving speech intelligibility can make communication more effective without requiring continuous transmission of all surrounding audio.

The low-power approach also makes the technology attractive for always-on applications, where conventional processing could place significant demands on a small battery.

Testing the VAD90 Prototype

To explore the technology in practice, IP Exchange received a VAD90 evaluation kit featuring voice activity detection and voice extraction functionality. The demonstration used a microphone, a laptop for audio recording, a noise source and a Nordic Semiconductor Power Profiler Kit II to observe the board’s power consumption.

The test began in bypass mode, allowing both speech and background noise to pass through to the recording. With AI mute enabled, the VAD90 detected when the presenter was speaking and muted the audio when no voice was present.

Voice extraction was then tested separately. With background street noise playing, the system demonstrated its ability to isolate speech from the surrounding audio, producing a cleaner voice signal at the output.

Finally, the power profiler was used to examine consumption while voice extraction remained active. The demonstration showed very low power consumption during operation, supporting the potential of the technology for power-constrained applications. However, the transcript does not provide a numerical measurement, so a precise power figure cannot be reported from this test.

Bringing Audio AI Closer to the Edge

The VAD90 demonstration highlights the practical potential of combining voice detection and extraction with ultra-low-power processing. For developers working on wearables, hearing assistance, smart microphones and other edge devices, NeuroVoice offers an interesting approach to implementing always-on audio intelligence without relying on cloud processing.

To learn more about NeuroVoice or request an evaluation kit, visit Polyn Technology’s website.