HyVox: Rethinking Low-Power Voice Intelligence towards Neuromorphic-von Neumann Edge Computing
| Source: arXiv AI
Tags: HyVox, neuromorphic computing, spiking neural networks, speaker identification, edge AI, Common Voice
HyVox splits speaker identification between a neuromorphic chip and a microcontroller, enrolling new users without weight updates and reaching 99.6% top-1 accuracy for two-person households at 1.63 mJ per inference.
Details
HyVox targets open-set speaker identification on tight energy budgets. A neuromorphic processor turns speech into a 32-spike-count embedding using sparse, event-driven compute, and a von Neumann microcontroller stores the speaker gallery, scores by cosine similarity, rejects unknown speakers and enrolls new users without retraining. The spiking backbone learns a time constant per neuron across fast, medium and slow timescales and was compressed 6.5x to 29.0k synapses to fit the chip. Chip-accurate spiking training, run through a differentiable replica of the chip's integer neuron update, doubled top-1 accuracy of the mapped backbone on unseen speakers. On 20 Common Voice households, HyVox hit 99.6% top-1 with 2 members and 94.9% with 20. Each 3-second window takes 0.96 s and 1.63 mJ of dynamic energy.