Following the launch of the VocalFusion® XVF3620 Voice Processor, XMOS has released a series of video demonstrations giving embedded engineers, system architects, and technical product managers an inside look at the silicon and how it tackles real-world problems.
The best hardware tests are the ones that step out of the lab and tackle the real world. We took the chip to the streets to test the real-world capabilities along with its robustness.
Live Test 1: AI Denoising in Negative SNR
In our first live test, we took the evaluation board onto a busy city street to evaluate speech intelligibility:
- AI Denoising in Negative Signal-to-Noise Ratios (SNR), even when background noise is physically louder than the speaker, voice clarity is achieved.
- Far-field capture, demonstrating clear voice recognition from over 5 meters away.
Watch as we toggle between the raw camera audio and the live XVF3620 output; demonstrating how the on-chip deep neural network separates human voice from non-stationary traffic chaos in real time.
Live Test 2: Acoustic Echo Cancellation & Adaptive Beamforming
Our second live test focuses on spatial isolation and full-duplex communication:
- Acoustic Echo Cancellation (AEC), We test voice barge-in by blasting loud audio right next to the microphones and evaluating the clear microphone output of the user.
- Adaptive Beamforming, using a two-microphone linear array, the chip dynamically steers digital blind spots to suppress off-axis interference by 3 to 6 dB.
