A 40-nm 3.9mW, 200words/Min Neural Signal Processor in Speech Decoding for Brain-Machine Interface

Area of Science:

• Neuroscience
• Computer Engineering
• Biomedical Engineering

Background:

• Brain-machine interface (BMI) technology facilitates direct communication between the human brain and external devices.
• Real-time speech decoding is a critical application for BMI, enabling communication for individuals with severe motor impairments.

Purpose of the Study:

• To develop an efficient neural signal processor for real-time speech decoding in brain-machine interfaces.
• To optimize energy consumption, memory footprint, and processing latency for neural signal processing.

Main Methods:

• Implemented speech attempt detection to reduce energy consumption and channel requirements.
• Utilized sparse encoding and mixed-precision arithmetic for weight encoding, reducing memory size.
• Employed computation reordering, partial sum caching, and exploitation of input/weight sparsity in the processing element (PE) array to decrease latency.
• Introduced an approximate top-k selection architecture in the beam search engine.

Main Results:

• Achieved a 46% reduction in energy consumption and decreased channels from 128 to 16 for speech attempt detection.
• Reduced neural network off-chip memory size by 80% and processing latency by 55% through optimized encoding and computation.
• PE array sparsity exploitation lowered latency by 95%, and mixed-precision multipliers reduced area by 27%.
• Attained a phone error rate of 16.6% and a word error rate of 23.5% for speech decoding.
• Demonstrated a maximum communication rate of 200 words/min, 16.7-to-42.6x faster than state-of-the-art, decoding up to 125,000 words.

Conclusions:

• The proposed neural signal processor significantly advances real-time BMI capabilities for speech decoding.
• The optimizations in energy, memory, and latency enable unprecedented decoding speeds and capacity.
• This technology holds promise for improving communication accessibility for individuals with speech and motor impairments.